From a960c8f91b1d0b47b6339628b751e15fcb9112bc Mon Sep 17 00:00:00 2001
From: Claude <claude@gitdan.com>
Date: Mon, 7 Sep 2026 00:20:45 -0500
Subject: [PATCH 1/4] refactor(cache): name the mutable set once, and measure
 it
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The set of paths a hardlink clone has to real-copy — dep-info,
build-script metadata, linked outputs, the pruned directories that hold
them — was spelled out inline in unshare_mutable_paths, in four find
invocations. Nothing else needed it, so one spelling was enough.

Something else needs it now: the prune has to know what a clone will
cost before it happens, and a sizer with its own copy of the predicates
would drift from the copier silently and in the dangerous direction — an
under-measured clone is one that starts and runs out of disk halfway
through unsharing. So the directory names become one array and the file
rules one dispatcher, applied by a callback per side, with each rule's
rationale moved to the rule rather than left at the old call site.

mutable_set_kb measures that set off a snapshot, skipping the subtrees
already measured whole so nothing is counted twice; clone_headroom_kb
scales it by a hand-written margin and floor for what the measurement
cannot see (cp -al materialising every directory for real, and the
unshare holding one subtree twice at its peak). Both residuals are named
where the function is, in both directions.

seed_source_candidates moves the seed's source-preference list into
cache-lib for the same reason: the prune ahead of it has to resolve the
same source the seed will clone, and two agreeing derivations are one
edit away from disagreeing.

No behaviour change — the copier applies the same rules to the same
tree, verified by the hardlink-clone suite's inode partition in both
directions.

Refs #20.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JXMQCJ5Eg5f9G9cfYzyh4Z
---
 scripts/cache-lib.sh       | 227 ++++++++++++++++++++++++++++++-------
 scripts/seed-target-dir.sh |   9 +-
 2 files changed, 192 insertions(+), 44 deletions(-)

diff --git a/scripts/cache-lib.sh b/scripts/cache-lib.sh
index d24ac4c..d195a5e 100755
--- a/scripts/cache-lib.sh
+++ b/scripts/cache-lib.sh
@@ -330,6 +330,62 @@ _holds_compiled_artifact() {
   return 1
 }
 
+# The directory names that select a mutable subtree, as one find predicate.
+#
+# Named once because THREE readers have to agree on it: the selection in
+# _mutable_dirs, the prune that skips those subtrees when it sizes the file
+# rules, and anything later that measures what a clone will cost. Two
+# spellings of this list would size a different tree than the one copied, and
+# the direction that fails is silent — an under-measured clone runs out of
+# disk mid-unshare, which is gitdan-actions#20.
+_MUTABLE_DIR_NAMES=( -name .fingerprint -o -name fingerprint -o -name run -o -name out )
+
+# _mutable_file_rules <fn>
+#
+# The FILE half of the mutable set, applied one rule at a time:
+#
+#     <fn> <label> <maxdepth|-> <find-predicate...>
+#
+# Same reason as the array above — `unshare_mutable_paths` copies these and
+# `mutable_set_kb` measures them, and a rule that exists in only one of the
+# two is exactly the under-estimate the headroom gate cannot survive. The
+# maxdepth is a separate field because GNU find wants it ahead of every other
+# predicate, so it cannot live inside the predicate vector.
+#
+# `-type f` is each caller's to add: the sizer needs it inside the `-o`
+# alternation it builds, the copier ahead of it.
+_mutable_file_rules() {
+  local fn="$1"
+  "$fn" 'dep-info files' - -name '*.d' || return 1
+  # Layout v1's build-script run metadata, which v2 groups under `run/` and v1
+  # leaves loose in the run unit's directory. `invoked.timestamp` is empty and
+  # carries its meaning in its mtime, which a shared inode carries too.
+  "$fn" 'build-script run metadata' - \
+    \( -name output -o -name root-output -o -name stderr -o -name invoked.timestamp \) || return 1
+  # Linked outputs. Unlike an rlib or an rmeta — which rustc writes to a
+  # temporary and renames into place — an executable or shared object is
+  # written by the LINKER, and the linker writes THROUGH an existing inode.
+  # Measured 2026-08-27 on cargo 1.93.1 stable, 1.96.0-nightly, 1.98.0-nightly
+  # (layout v1) and 1.100.0-nightly (e8cb624d5, layout v2), mold and the
+  # default linker alike: a `cargo test --no-run` binary in a `cp -al` clone
+  # rewrote the SOURCE's copy of itself in place, under both layouts.
+  #
+  # What separates that from the executables measured INTACT is not
+  # established. Every intact case observed was one Cargo has to re-create
+  # anyway to maintain an uplift hardlink — a bin target's
+  # `deps/<bin>-<hash>`, twinned at `<profile>/<bin>`. Whether the twin is the
+  # mechanism or a correlate of it was not determined, and the rule below does
+  # not depend on the answer: exempting twinned executables would recover no
+  # bytes this function newly copies. See gitdan-actions#17.
+  #
+  # The executable bit is the discriminator because it is the linker's own
+  # output that is at risk, not the directory it happens to land in — `.rlib`,
+  # `.rmeta` and `incremental/` stay shared and they are the bytes that matter.
+  "$fn" 'linked outputs' - -perm -u+x || return 1
+  "$fn" '.rustc_info.json' 3 -name '.rustc_info.json' || return 1
+  return 0
+}
+
 # The directories `unshare_mutable_paths` replaces, under either layout.
 #
 # All four names are pruned, so nothing selected here can contain anything else
@@ -380,10 +436,24 @@ _mutable_dirs() {
     fi
     printf '%s\n' "$d"
   done < <(find "$root" -type d \
-    \( -name .fingerprint -o -name fingerprint -o -name run -o -name out \) \
+    \( "${_MUTABLE_DIR_NAMES[@]}" \) \
     -prune -print 2>/dev/null)
 }
 
+# _mutable_file_rules' callback for the copying side. The root travels in a
+# variable rather than an argument because the callback's own signature is the
+# rule's, and every rule has to reach the same tree.
+_unshare_one_rule() {
+  local label="$1" maxdepth="$2"; shift 2
+  local -a depth=()
+  [ "$maxdepth" = - ] || depth=(-maxdepth "$maxdepth")
+  _unshare_files "$_MUTABLE_ROOT" ${depth[@]+"${depth[@]}"} -type f "$@" || {
+    echo "::error::unshare_mutable_paths: failed to unshare ${label} under ${_MUTABLE_ROOT}" >&2
+    return 1
+  }
+  return 0
+}
+
 # THE load-bearing function of this whole design.
 #
 # A hardlink clone is only safe if every write the clone's build performs
@@ -493,6 +563,7 @@ _mutable_dirs() {
 unshare_mutable_paths() {
   local root="$1" d
   [ -d "$root" ] || return 0
+  _MUTABLE_ROOT="$root"
   # The list is materialised in full before anything is replaced: each
   # replacement deletes and recreates a directory, and a live `find` walk over
   # a tree being mutated underneath it is a needless hazard.
@@ -505,45 +576,121 @@ unshare_mutable_paths() {
       return 1
     }
   done
-  _unshare_files "$root" -type f -name '*.d' || {
-    echo "::error::unshare_mutable_paths: failed to unshare dep-info files under ${root}" >&2
-    return 1
-  }
-  # Layout v1's build-script run metadata, which v2 groups under `run/` and v1
-  # leaves loose in the run unit's directory. `invoked.timestamp` is empty and
-  # carries its meaning in its mtime, which a shared inode carries too.
-  _unshare_files "$root" -type f \
-    \( -name output -o -name root-output -o -name stderr -o -name invoked.timestamp \) || {
-    echo "::error::unshare_mutable_paths: failed to unshare build-script run metadata under ${root}" >&2
-    return 1
-  }
-  # Linked outputs. Unlike an rlib or an rmeta — which rustc writes to a
-  # temporary and renames into place — an executable or shared object is
-  # written by the LINKER, and the linker writes THROUGH an existing inode.
-  # Measured 2026-08-27 on cargo 1.93.1 stable, 1.96.0-nightly, 1.98.0-nightly
-  # (layout v1) and 1.100.0-nightly (e8cb624d5, layout v2), mold and the
-  # default linker alike: a `cargo test --no-run` binary in a `cp -al` clone
-  # rewrote the SOURCE's copy of itself in place, under both layouts.
-  #
-  # What separates that from the executables measured INTACT is not
-  # established. Every intact case observed was one Cargo has to re-create
-  # anyway to maintain an uplift hardlink — a bin target's
-  # `deps/<bin>-<hash>`, twinned at `<profile>/<bin>`. Whether the twin is the
-  # mechanism or a correlate of it was not determined, and the rule below does
-  # not depend on the answer: exempting twinned executables would recover no
-  # bytes this function newly copies. See gitdan-actions#17.
-  #
-  # The executable bit is the discriminator because it is the linker's own
-  # output that is at risk, not the directory it happens to land in — `.rlib`,
-  # `.rmeta` and `incremental/` stay shared and they are the bytes that matter.
-  _unshare_files "$root" -type f -perm -u+x || {
-    echo "::error::unshare_mutable_paths: failed to unshare linked outputs under ${root}" >&2
-    return 1
-  }
-  _unshare_files "$root" -maxdepth 3 -type f -name '.rustc_info.json' || {
-    echo "::error::unshare_mutable_paths: failed to unshare .rustc_info.json under ${root}" >&2
-    return 1
-  }
+  _mutable_file_rules _unshare_one_rule || return 1
+  return 0
+}
+
+# ---------------------------------------------------------------------------
+# What the seed will clone, and what that clone costs in disk
+# ---------------------------------------------------------------------------
+
+# The sources seed-target-dir.sh considers, most specific first, as
+# `<dir>:<label>` lines.
+#
+# Read by the seed, which clones the first one that exists, and by the prune
+# that has to size the volume for that clone BEFORE it happens. One derivation
+# rather than two agreeing ones: a prune that sizes a different tree than the
+# seed clones is measuring nothing, and nothing downstream would say so.
+seed_source_candidates() {
+  local root="$1" own_key="$2" base_key="$3" fallback="${4:-}"
+  [ -n "$base_key" ] && printf '%s:base snapshot\n' "$(snapshot_dir_for "$root" "$base_key")"
+  printf '%s:own snapshot\n' "$(snapshot_dir_for "$root" "$own_key")"
+  [ -n "$fallback" ] && printf '%s:fallback dir\n' "$fallback"
+  return 0
+}
+
+# The directory the seed will actually hardlink-clone on this run, or nothing
+# at all when it will not clone: its own target dir already exists (the seed
+# reuses it and returns before the candidate list is consulted), or no
+# candidate exists (it starts cold).
+seed_clone_source() {
+  local root="$1" own_key="$2" base_key="$3" fallback="${4:-}" entry src
+  [ -d "$(target_dir_for "$root" "$own_key")" ] && return 0
+  while IFS= read -r entry; do
+    src="${entry%%:*}"
+    if [ -d "$src" ]; then printf '%s' "$src"; return 0; fi
+  done < <(seed_source_candidates "$root" "$own_key" "$base_key" "$fallback")
+  return 0
+}
+
+# _mutable_file_rules' callback for the measuring side.
+#
+# One find per rule, skipping the subtrees `_mutable_dirs` already selects
+# whole — those are measured by the `du` in mutable_set_kb, and counting a
+# file twice would inflate the requirement into evicting caches nothing
+# needed. `%k` is allocated 1K blocks, the same unit `du -sk` reports, so the
+# two halves add.
+_size_one_rule() {
+  local maxdepth="$2"; shift 2
+  local -a depth=()
+  [ "$maxdepth" = - ] || depth=(-maxdepth "$maxdepth")
+  find "$_MUTABLE_ROOT" ${depth[@]+"${depth[@]}"} \
+    \( "${_MUTABLE_DIR_NAMES[@]}" \) -prune -o \
+    -type f \( "$@" \) -printf '%k\n' 2>/dev/null
+  return 0
+}
+
+# The kilobytes `unshare_mutable_paths` will really-copy out of <dir> — the
+# part of a hardlink clone that costs new disk, as opposed to the `.rlib`,
+# `.rmeta` and `incremental/` bytes that stay shared with the source.
+#
+# Measured off the SAME enumeration the copier uses (`_mutable_dirs` and
+# `_mutable_file_rules`), which is the only thing that makes this a
+# measurement rather than an estimate.
+#
+# Two residuals, both named because neither is bounded away:
+#
+#   OVER by any file matching two rules at once — an executable named
+#   `output`, say. Rare, and small.
+#   UNDER by the `*.d` files inside an `out` directory that _mutable_dirs
+#   leaves SHARED (the compile-unit case: it holds an .rlib and has no
+#   build-script record beside it). Such a directory holds a library artifact
+#   by definition, so what is missed is dep-info, not executables. Also under
+#   by a `.rustc_info.json` deeper than the copier's own maxdepth, which is
+#   kilobytes.
+#
+# The margin in clone_headroom_kb is what covers the under-count; it is not
+# there to make the measurement optional.
+mutable_set_kb() {
+  local root="$1"
+  [ -d "$root" ] || { printf '0'; return 0; }
+  _MUTABLE_ROOT="$root"
+  {
+    _mutable_dirs "$root" | tr '\n' '\0' | xargs -0 -r du -sk 2>/dev/null | awk '{print $1}'
+    _mutable_file_rules _size_one_rule
+  } | awk '{s += $1} END { printf "%d", s + 0 }'
+  return 0
+}
+
+# How much free space the seed needs on the volume before it clones <dir>.
+#
+# CACHE_CLONE_HEADROOM_PERCENT scales the measured mutable set;
+# CACHE_CLONE_HEADROOM_FLOOR_KB is added on top. BOTH DEFAULTS ARE
+# HAND-WRITTEN — nothing measures them, and they are separate because they
+# cover different things:
+#
+#   The percentage covers what scales with the tree: `unshare_subtree` stages
+#   each mutable directory through a sibling copy before dropping the shared
+#   original, so at its peak one subtree is held twice, and the under-count
+#   named on mutable_set_kb scales with the tree too.
+#   The floor covers what does not: `cp -al` materialises every DIRECTORY for
+#   real (only files are linked), and a Bevy-sized target dir has hundreds of
+#   thousands of them.
+#
+# Both are overridable, and the direction of error is deliberate. Over-asking
+# evicts a cache that would have fitted, costing one branch a cold start;
+# under-asking lets the clone start and run out of disk halfway through the
+# unshare, which fails the job with an error naming a staging path — the
+# failure gitdan-actions#20 is filed about.
+CACHE_CLONE_HEADROOM_PERCENT="${CACHE_CLONE_HEADROOM_PERCENT:-150}"
+CACHE_CLONE_HEADROOM_FLOOR_KB="${CACHE_CLONE_HEADROOM_FLOOR_KB:-2097152}"
+
+clone_headroom_kb() {
+  local src="${1:-}" kb
+  if [ -z "$src" ] || [ ! -d "$src" ]; then printf '0'; return 0; fi
+  kb=$(mutable_set_kb "$src")
+  awk -v k="$kb" -v pct="$CACHE_CLONE_HEADROOM_PERCENT" -v floor="$CACHE_CLONE_HEADROOM_FLOOR_KB" \
+    'BEGIN { printf "%d", (k * pct / 100) + floor }'
   return 0
 }
 
diff --git a/scripts/seed-target-dir.sh b/scripts/seed-target-dir.sh
index 8451d10..bf51207 100755
--- a/scripts/seed-target-dir.sh
+++ b/scripts/seed-target-dir.sh
@@ -61,10 +61,11 @@ fi
 #      snapshot already exists.
 #   3. an explicit fallback directory — migration off a pre-existing flat
 #      cache, so the first run under this scheme isn't a needless cold build.
-CANDIDATES=()
-[ -n "$BASE_KEY" ] && CANDIDATES+=("$(snapshot_dir_for "$ROOT" "$BASE_KEY"):base snapshot")
-CANDIDATES+=("$(snapshot_dir_for "$ROOT" "$OWN_KEY"):own snapshot")
-[ -n "$FALLBACK" ] && CANDIDATES+=("${FALLBACK}:fallback dir")
+#
+# The list itself lives in cache-lib.sh because prune-cache.sh reads it too:
+# it runs ahead of this step and has to free enough disk for the clone below,
+# which means resolving the same source this loop will pick.
+mapfile -t CANDIDATES < <(seed_source_candidates "$ROOT" "$OWN_KEY" "$BASE_KEY" "$FALLBACK")
 
 for entry in "${CANDIDATES[@]}"; do
   src="${entry%%:*}"; label="${entry#*:}"

From 07ba53ca799ab1f7e5eef2abe5de9f4879dfa347 Mon Sep 17 00:00:00 2001
From: Claude <claude@gitdan.com>
Date: Mon, 7 Sep 2026 00:21:12 -0500
Subject: [PATCH 2/4] fix(prune): reclaim merged branches, and size the volume
 for the clone
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Two assumptions in the eviction pass did not hold on this forge, and
between them a volume filled up three times in three days with nothing
reclaimed automatically. Both are replaced here; the pass also moves
ahead of the seed, which is the only order in which its work can help
the run performing it.

LIVENESS. Pass 1 evicted a cache only when its branch was gone from
origin. Gitea keeps a PR's branch after the merge unless the repo opts
into delete-on-merge, and zemyna does not — so ls-remote reports fifty
merged branches and the signal fires for none of them. A second signal
is added beside it: a branch still on origin whose tip is an ancestor of
a protected branch's tip holds no commit that branch does not, so its
cache will never be read again and goes in the same unconditional pass.

Ancestry is answered from the commits in the job's own checkout, so the
answer "cannot tell" exists and stays distinct from "not merged" at both
granularities. A shallow checkout withholds the signal entirely, since a
missing object is its normal case rather than evidence. A single branch
whose tip is not in the checkout is kept, with a warning naming it. A
squash or rebase merge leaves no ancestry and reads as live until the
branch is deleted. All three are missed reclamations, which cost disk;
the other direction costs a branch its cache mid-build.

HEADROOM. Passes 2 and 3 gated on a percentage of the volume, which
cannot express the failure they have to prevent: a clone runs out of
disk while unsharing its mutable paths, and how much that needs is a
property of the snapshot rather than of the disk. Staging failed at 34 G
free and passed at 74 G, so a 10% floor — 19 G here — never fired first.
The requirement is now measured per run off the very source the seed
will read, the pass evicts oldest-first until it is met and stops there,
and falling short of it fails with the shortfall and every directory it
kept, rather than letting the seed fail seconds later against a staging
path that names none of that. min-free-percent survives as an additional
floor, defaulting to 0, and falling short of that one is still a warning
and a self-clear.

ORDERING. The prune step ran after the seed, so each run freed space for
the next one. It now runs between resolve and seed. Two things that
makes newly reachable are closed: the source about to be cloned is
excluded from every pass by name, and a concurrent job's target dir
already carries its lock from the instant it appears under its final
name, so nothing is seen unlocked that is in use.

Red-proven: sixteen assertions across the four new scenarios fail
against the pre-fix scripts, including the zemyna layout evicting
nothing where it should evict exactly one directory, and the headroom
scenario exiting 0 where it should exit 1.

Refs #20.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JXMQCJ5Eg5f9G9cfYzyh4Z
---
 cargo-cache/action.yml          |  73 ++++++--
 scripts/prune-cache-selftest.sh | 195 +++++++++++++++++++-
 scripts/prune-cache.sh          | 304 ++++++++++++++++++++++++++++----
 3 files changed, 520 insertions(+), 52 deletions(-)

diff --git a/cargo-cache/action.yml b/cargo-cache/action.yml
index 84d3777..4006673 100644
--- a/cargo-cache/action.yml
+++ b/cargo-cache/action.yml
@@ -29,9 +29,18 @@ inputs:
     required: false
     default: 'dev main'
   min-free-percent:
-    description: 'Prune when free space on the cache volume drops below this percentage.'
+    description: >-
+      An ADDITIONAL free-space floor, as a percentage of the cache volume.
+      The prune's own requirement is derived per run from what the seed is
+      about to clone — the mutable set it has to real-copy out of the source
+      snapshot, which is a property of that snapshot and not of the volume —
+      and this floor only ever raises it. 0, the default, leaves the derived
+      requirement as the only gate. Set it to keep headroom for something
+      other than the clone (the build's own output, another job on the same
+      volume); it will not make the clone fit, because it does not know how
+      big the clone is.
     required: false
-    default: '10'
+    default: '0'
   restore-mtimes:
     description: >-
       Restore every tracked file's mtime from git history. Requires a
@@ -40,13 +49,20 @@ inputs:
     required: false
     default: 'true'
   prune:
-    description: 'Run the eviction pass (dead-branch liveness + disk pressure).'
+    description: >-
+      Run the eviction pass (dead-branch liveness + disk pressure). It runs
+      BEFORE the seed step, so what it frees is available to the clone that
+      step makes.
     required: false
     default: 'true'
   liveness-prune:
     description: >-
-      Within the prune pass, remove caches for branches that no longer exist
-      on origin. Set to false on a runner that cannot reach origin.
+      Within the prune pass, remove caches for branches that are dead: gone
+      from origin, or still on origin with a tip already merged into a
+      protected branch. The second signal is what reclaims anything at all on
+      a forge that keeps branches after merge, and it needs the protected
+      branches' commits in the checkout — a shallow one withholds it and says
+      so. Set to false on a runner that cannot reach origin.
     required: false
     default: 'true'
   own-ref:
@@ -169,6 +185,39 @@ runs:
           echo "CI_WATERMARK_FILE=${WATERMARK}"
         } >> "$GITHUB_ENV"
 
+    # Runs BEFORE the seed, which is the only order in which its work can
+    # help: the eviction it performs is what makes room for the clone the seed
+    # step is about to make, and the requirement it evicts against is measured
+    # off the snapshot that clone will read. Running afterwards — where this
+    # step used to be — meant every run freed space for the NEXT one and the
+    # seed met whatever the last run happened to leave.
+    #
+    # Two things this ordering has to be safe against, and is:
+    #
+    #   The source it is about to read is excluded from every pass by name
+    #   (see protected_reason in prune-cache.sh), so pass 1 cannot take the
+    #   snapshot out from under the seed that follows it.
+    #   A concurrent job's target dir carries its lock from the instant it
+    #   appears under its final name — the seed writes it into the staging
+    #   tree before the rename — so there is no window in which running this
+    #   earlier sees an unlocked directory somebody is using.
+    - if: ${{ inputs.prune == 'true' }}
+      shell: bash
+      env:
+        STALE_LOCK_SECONDS: ${{ inputs.stale-lock-seconds }}
+        CACHE_LIVENESS: ${{ inputs.liveness-prune }}
+      run: |
+        set -euo pipefail
+        SCRIPTS=$(cd "${{ github.action_path }}/.." && pwd)/scripts
+        bash "${SCRIPTS}/prune-cache.sh" \
+          "${{ steps.resolve.outputs.cache-root }}" \
+          "${{ steps.resolve.outputs.target-dir }}" \
+          "${{ inputs.protected-branches }}" \
+          "${{ inputs.min-free-percent }}" \
+          "${{ steps.resolve.outputs.cache-key }}" \
+          "${{ steps.resolve.outputs.base-key }}" \
+          "${{ inputs.seed-fallback-dir }}"
+
     # Seeds this ref's target dir from the base's published snapshot. See
     # scripts/seed-target-dir.sh — the staging-then-atomic-rename is what
     # makes concurrent jobs sharing one cache key safe by construction rather
@@ -228,17 +277,3 @@ runs:
         CARGO_TARGET_DIR: ${{ steps.resolve.outputs.target-dir }}
         CI_WATERMARK_FILE: ${{ steps.resolve.outputs.watermark-file }}
       run: bash "$(cd "${{ github.action_path }}/.." && pwd)/scripts/restore-mtimes.sh"
-
-    - if: ${{ inputs.prune == 'true' }}
-      shell: bash
-      env:
-        STALE_LOCK_SECONDS: ${{ inputs.stale-lock-seconds }}
-        CACHE_LIVENESS: ${{ inputs.liveness-prune }}
-      run: |
-        set -euo pipefail
-        SCRIPTS=$(cd "${{ github.action_path }}/.." && pwd)/scripts
-        bash "${SCRIPTS}/prune-cache.sh" \
-          "${{ steps.resolve.outputs.cache-root }}" \
-          "${{ steps.resolve.outputs.target-dir }}" \
-          "${{ inputs.protected-branches }}" \
-          "${{ inputs.min-free-percent }}"
diff --git a/scripts/prune-cache-selftest.sh b/scripts/prune-cache-selftest.sh
index ab7c8cf..ff5cca8 100755
--- a/scripts/prune-cache-selftest.sh
+++ b/scripts/prune-cache-selftest.sh
@@ -46,10 +46,36 @@
 #      empty a tree its owner may still restore under a live cache name. Its
 #      fixture is an OLD directory renamed a moment ago — production's shape,
 #      and what lets it tell the two timestamps apart.
+#  15. A MERGED-BUT-UNDELETED BRANCH IS DEAD TOO. This forge keeps branches
+#      after merge, so `ls-remote` reports them forever and scenario 1's
+#      signal never fires for them — which is how three 40 GB caches sat on a
+#      full volume until somebody removed them by hand (gitdan-actions#20). A
+#      branch whose tip is an ancestor of a protected branch's tip is pruned
+#      like a deleted one; an unmerged branch beside it is not.
+#  16. AND "CANNOT TELL" IS STILL NOT DEATH, at both granularities: a branch
+#      whose tip is not in this checkout is kept with a warning naming it,
+#      and a shallow checkout — where a missing object is the normal case —
+#      withholds the whole signal rather than reading it as "nothing merged".
+#      The deleted-branch signal keeps working in both.
+#  17. THE FREE-SPACE REQUIREMENT IS MEASURED OFF THE SOURCE, not taken as a
+#      percentage of the volume: the pass evicts until the clone the seed is
+#      about to make fits, and stops there rather than draining the volume.
+#      When it cannot get there it FAILS, naming the shortfall and every
+#      directory it kept instead — because the seed would otherwise fail
+#      seconds later against a staging path that names nothing.
+#  18. AND ON THE LAYOUT THAT PRODUCED THE BUG: three equal-sized caches, one
+#      of them a merged-but-undeleted branch's, with disk to spare. Exactly
+#      that one goes. Equal sizes and no pressure are the point — nothing but
+#      the merge state can be what decides.
 set -euo pipefail
 script_dir=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
 . "$script_dir/cache-lib.sh"
 prune="$script_dir/prune-cache.sh"
+# Two scenarios below put a stub of a real tool on PATH for one command.
+# Captured once, here, rather than read back at each of those sites: a `$PATH`
+# read after the first of them is indistinguishable, to a static check, from
+# reading the modification the subshell lost.
+outer_path="$PATH"
 
 scratch=$(mktemp -d)
 trap 'rm -rf "$scratch"' EXIT
@@ -235,7 +261,7 @@ done
 exec "$real_du" "\$@"
 EOF
 chmod +x "$scratch/bin/du"
-( PATH="$scratch/bin:$PATH"; run_prune "1000000 900000" )
+( PATH="$scratch/bin:$outer_path"; run_prune "1000000 900000" )
 [ -e "$root/.reading-target-$DEAD-racer" ] || fail "the racing marker was never published — scenario 12 proves nothing"
 assert_kept "$root/target-$DEAD" "a cache claimed inside the eviction window is not unlinked"
 assert_kept "$root/target-$DEAD/blob" "the reprieved cache still has its contents"
@@ -286,5 +312,172 @@ assert_kept "$aside" "an aside younger than the settle window is not reclaimed"
 assert_kept "$aside/blob" "and is left intact, not part-way emptied"
 assert_log "may still be evicting it" "the deferral gives its actual reason"
 
+echo
+echo "=== 15: a merged-but-undeleted branch is dead too ==="
+# This forge keeps a PR's branch after the merge, so `ls-remote` reports it
+# forever. Built the way that happens: a branch merged into dev with a merge
+# commit, still pushed, beside one branched at the same point and NOT merged.
+git="git -c user.email=t@t -c user.name=t -c commit.gpgsign=false"
+(
+  cd "$work"
+  git checkout -q dev
+  git checkout -q -b feat/merged
+  $git commit -q --allow-empty -m merged
+  git checkout -q dev
+  $git merge -q --no-ff feat/merged -m "merge feat/merged"
+  git checkout -q -b feat/unmerged
+  $git commit -q --allow-empty -m unmerged
+  git checkout -q dev
+  git push -q origin dev feat/merged feat/unmerged
+)
+MERGED=$(cache_key feat/merged); UNMERGED=$(cache_key feat/unmerged)
+reset_cache
+mk "target-$MERGED"   '2030-01-01'
+mk "snapshot-$MERGED" '2030-01-01'
+mk "target-$UNMERGED" '2020-01-01'   # older, deliberately: merge state decides, not age
+run_prune "1000000 900000"           # 90% free: no pressure at all
+assert_log "merged-branch detection anchored on" "the pass says what it anchored ancestry on"
+assert_gone "$root/target-$MERGED" "a merged branch's cache is pruned though its branch is still on origin"
+assert_gone "$root/snapshot-$MERGED" "and so is its snapshot"
+assert_kept "$root/target-$UNMERGED" "an unmerged branch's cache survives, though it is the older of the two"
+assert_log "merged into dev" "the eviction names the branch it was merged into"
+
+echo
+echo "=== 16: 'cannot tell' is not death, per branch and per checkout ==="
+# A branch whose tip this checkout has never seen. Pushed from a second clone,
+# so `ls-remote` reports a SHA that `$work` holds no object for — which is
+# what "cannot determine" actually looks like, rather than a stubbed failure.
+other="$scratch/other"; git clone -q "$origin" "$other"
+(
+  cd "$other"
+  git checkout -q -b feat/elsewhere origin/dev
+  git -c user.email=t@t -c user.name=t -c commit.gpgsign=false commit -q --allow-empty -m elsewhere
+  git push -q origin feat/elsewhere
+)
+ELSEWHERE=$(cache_key feat/elsewhere)
+reset_cache
+mk "target-$ELSEWHERE" '2030-01-01'
+mk "target-$MERGED"    '2030-01-01'
+run_prune "1000000 900000"
+assert_kept "$root/target-$ELSEWHERE" "a branch whose tip is not in this checkout is kept, not classified dead"
+assert_log "cannot tell merged from live" "and the undecidable branch is named, not silently skipped"
+assert_gone "$root/target-$MERGED" "while a branch it CAN decide is still pruned in the same pass"
+
+# A shallow checkout, where a missing object is the ordinary case rather than
+# a signal — so the whole merged half is withheld. The deleted-branch half is
+# unaffected, which is what keeps this a narrowing rather than an outage.
+shallow="$scratch/shallow"; git clone -q --depth 1 -b dev "file://$origin" "$shallow"
+[ "$(git -C "$shallow" rev-parse --is-shallow-repository)" = true ] \
+  || fail "the fixture clone is not shallow — scenario 16's second half proves nothing"
+reset_cache
+mk "target-$MERGED" '2030-01-01'
+(
+  cd "$shallow"
+  CACHE_DF_OVERRIDE="1000000 900000" GITHUB_STEP_SUMMARY="$scratch/summary" \
+    bash "$prune" "$root" "$root/target-$OWN" "dev main" 10
+) > "$scratch/log" 2>&1 || { cat "$scratch/log"; fail "prune-cache.sh exited non-zero in a shallow checkout"; }
+assert_log "checkout is shallow" "a shallow checkout withholds the merged signal and says why"
+assert_kept "$root/target-$MERGED" "and keeps a merged branch's cache rather than guessing"
+assert_gone "$root/target-$DEAD" "while the deleted-branch signal still fires"
+
+echo
+echo "=== 17: the free-space requirement is measured off the clone's source ==="
+# A `df` that answers from the cache root's actual size, because the property
+# under test is that the pass STOPS once the requirement is met — which a
+# fixed CACHE_DF_OVERRIDE cannot express, since evicting never changes it.
+mkdir -p "$scratch/bin17"
+real_du=$(command -v du)
+build_seed_fixture() {
+  rm -rf "$root"; mkdir -p "$root"
+  # The source the seed is about to clone. 8 MB of dep-info, which
+  # unshare_mutable_paths has to real-copy, beside 16 MB of .rlib that it
+  # leaves hardlinked — so a requirement derived from the SIZE of the source
+  # would be three times the one derived from its mutable set.
+  mkdir -p "$root/snapshot-$DEV/debug/.fingerprint/unit" "$root/snapshot-$DEV/debug/deps"
+  head -c $((8 * 1024 * 1024)) /dev/zero > "$root/snapshot-$DEV/debug/.fingerprint/unit/dep-lib"
+  head -c $((16 * 1024 * 1024)) /dev/zero > "$root/snapshot-$DEV/debug/deps/libx.rlib"
+  touch -d '2020-01-01' "$root/snapshot-$DEV/.cache-last-used"
+  # Three live, unmerged branches' caches of 4 MB each, oldest first.
+  local i=0
+  for b in a b c; do
+    i=$((i + 1))
+    mkdir -p "$root/target-$(cache_key "feat/$b")"
+    head -c $((4 * 1024 * 1024)) /dev/zero > "$root/target-$(cache_key "feat/$b")/blob"
+    touch -d "202${i}-01-01" "$root/target-$(cache_key "feat/$b")/.cache-last-used"
+  done
+  # A volume with 2 MB to spare: under the requirement, over nothing else.
+  cap=$(( $($real_du -sk "$root" | awk '{print $1}') + 2048 ))
+  cat > "$scratch/bin17/df" <<DFEOF
+#!/usr/bin/env bash
+used=\$($real_du -sk "$root" | awk '{print \$1}')
+echo "Filesystem 1024-blocks Used Available Capacity Mounted-on"
+echo "fake $cap \$used \$(( $cap - used )) 50% $root"
+DFEOF
+  chmod +x "$scratch/bin17/df"
+}
+(
+  cd "$work"
+  for b in a b c; do
+    git checkout -q dev
+    git checkout -q -b "feat/$b"
+    git -c user.email=t@t -c user.name=t -c commit.gpgsign=false commit -q --allow-empty -m "$b"
+  done
+  git checkout -q dev
+  git push -q origin feat/a feat/b feat/c
+)
+# run_seeded_prune <own-ref> <base-ref> — the form cargo-cache/action.yml uses:
+# the same pass, told what the seed step it now runs ahead of will clone. The
+# headroom knobs are pinned so the arithmetic is the fixture's, not the
+# defaults' (whose 2 GiB floor would dwarf any fixture on a test host).
+seeded_rc=0
+run_seeded_prune() {
+  seeded_rc=0
+  PATH="$scratch/bin17:$outer_path" \
+    CACHE_CLONE_HEADROOM_PERCENT=100 CACHE_CLONE_HEADROOM_FLOOR_KB=1024 \
+    GITHUB_STEP_SUMMARY="$scratch/summary" \
+    bash "$prune" "$root" "$root/target-$(cache_key "$1")" "dev main" 0 \
+      "$(cache_key "$1")" "$(cache_key "$2")" "" \
+    > "$scratch/log" 2>&1 || seeded_rc=$?
+}
+
+build_seed_fixture
+run_seeded_prune feat/own dev
+[ "$seeded_rc" = 0 ] || { cat "$scratch/log"; fail "prune-cache.sh exited ${seeded_rc} with the requirement satisfiable"; }
+assert_log "measured from its mutable set" "the requirement says where it came from"
+assert_gone "$root/target-$(cache_key feat/a)" "the oldest cache is evicted to make room for the clone"
+assert_gone "$root/target-$(cache_key feat/b)" "and the next oldest, because one was not enough"
+assert_kept "$root/target-$(cache_key feat/c)" "and the pass STOPS there rather than draining the volume"
+assert_kept "$root/snapshot-$DEV" "the source the seed is about to clone is never a candidate"
+
+# Nothing eligible: every sibling is held open by a running job. The pass
+# cannot reach the requirement, and the seed that follows would fail against a
+# staging path naming none of this.
+build_seed_fixture
+for b in a b c; do date +%s > "$root/target-$(cache_key "feat/$b")/.ci-lock-ci-1"; done
+run_seeded_prune feat/own dev
+[ "$seeded_rc" = 1 ] || { cat "$scratch/log"; fail "expected exit 1 when the clone cannot fit, got ${seeded_rc}" ; }
+ok "a clone that cannot be made to fit fails the pass rather than the seed"
+assert_log "short by" "the failure names the shortfall"
+assert_log "held open by a running job" "and what was kept instead of it, with the reason"
+assert_kept "$root/target-$(cache_key feat/a)" "a locked cache is still not evicted, however tight the disk"
+
+echo
+echo "=== 18: the layout that produced the bug ==="
+# zemyna's volume on 2026-09-07: the base branch's snapshot and target dir,
+# plus one target dir for a branch merged the day before and never deleted.
+# Equal sizes and 90% free, so neither age nor pressure nor size can be what
+# decides — only the merge state.
+rm -rf "$root"; mkdir -p "$root"
+mk "snapshot-$DEV"  '2026-09-01'
+mk "target-$DEV"    '2026-09-01'
+mk "target-$MERGED" '2026-09-06'
+run_prune "1000000 900000"
+assert_kept "$root/snapshot-$DEV" "the base snapshot stays"
+assert_kept "$root/target-$DEV" "and the base target dir stays"
+assert_gone "$root/target-$MERGED" "and the merged-but-undeleted branch's cache is the one reclaimed"
+[ "$(grep -c 'pruned dead-branch cache' "$scratch/log")" = 1 ] \
+  || { cat "$scratch/log"; fail "expected exactly one eviction on the zemyna layout"; }
+ok "exactly one directory is evicted, and it is that one"
+
 echo
 echo "prune-cache-selftest: ${pass_count} assertions passed"
diff --git a/scripts/prune-cache.sh b/scripts/prune-cache.sh
index 6efa4fc..7e1827c 100755
--- a/scripts/prune-cache.sh
+++ b/scripts/prune-cache.sh
@@ -1,8 +1,16 @@
 #!/usr/bin/env bash
 # Eviction for the per-ref cache directories on the persistent volume.
 #
-# Usage: prune-cache.sh <cache-root> <own-target-dir> <protected-branches> <min-free-percent>
+# Usage: prune-cache.sh <cache-root> <own-target-dir> <protected-branches> \
+#          <min-free-percent> [own-key] [base-key] [fallback-dir]
 #   protected-branches  space-separated raw refs (e.g. "dev main")
+#   min-free-percent    a FLOOR, not the gate — 0 to rely on the derived
+#                       requirement alone (the default)
+#   own-key, base-key, fallback-dir
+#                       the same three the seed step resolves its source from.
+#                       Given them, this pass sizes the volume for the clone
+#                       that step is about to make; without them it has only
+#                       the percentage floor, and says so.
 #
 # Optional environment:
 #   STALE_LOCK_SECONDS  age past which a .ci-lock-* marker is treated as
@@ -16,19 +24,41 @@
 #
 # Three passes, in order:
 #
-#   1. LIVENESS — every target-*/snapshot-* directory whose branch no longer
-#      exists on origin is removed UNCONDITIONALLY, not gated on free space.
-#      A directory for a branch deleted days ago is pure loss: nothing will
-#      ever read it again, since a merged PR's branch cannot be reopened.
-#      Waiting for disk pressure to notice means paying for it until then.
-#      Skipped entirely, loudly, if the liveness signal itself is
-#      unavailable — "couldn't determine" is never folded into "dead".
-#   2. PRESSURE — if free space is still under the threshold, evict remaining
-#      (now necessarily live) directories oldest-first until it clears.
-#   3. SELF-CLEAR — if pass 2 still isn't enough, wipe this run's own target
-#      dir and pay a cold rebuild, reported to the job summary as well as the
-#      log, because a warning on a green run is what lets a silently 4x-slower
-#      job go unnoticed.
+#   1. LIVENESS — every target-*/snapshot-* directory whose branch is DEAD is
+#      removed UNCONDITIONALLY, not gated on free space. A directory for a
+#      branch nothing will build again is pure loss; waiting for disk pressure
+#      to notice means paying for it until then. Two signals make a branch
+#      dead, and the second exists because the first alone is inert on a forge
+#      that keeps branches after merge (gitdan-actions#20):
+#
+#        DELETED — the branch is no longer on origin at all.
+#        MERGED  — the branch is still on origin, but its tip is an ancestor
+#                  of a protected branch's tip, so every commit it holds is
+#                  already on the branch its cache would be re-cloned from.
+#
+#      Skipped entirely, loudly, if the signal itself is unavailable —
+#      "couldn't determine" is never folded into "dead", for either signal
+#      and at either granularity: a checkout that cannot answer ancestry at
+#      all skips the merged half, and a single branch whose tip is not in the
+#      checkout is kept with a warning naming it.
+#   2. PRESSURE — if free space is under the requirement, evict remaining
+#      (now necessarily live) directories oldest-first until it clears. The
+#      requirement is what the seed step is about to need to clone its source,
+#      MEASURED off that source (see clone_headroom_kb in cache-lib.sh), and a
+#      percentage floor only if one is configured. A percentage cannot express
+#      this: the failure it has to prevent is a clone running out of disk
+#      part-way through unsharing its mutable paths, and how much that needs
+#      is a property of the snapshot, not of the volume.
+#   3. SELF-CLEAR — if pass 2 still isn't enough for the percentage floor,
+#      wipe this run's own target dir and pay a cold rebuild, reported to the
+#      job summary as well as the log, because a warning on a green run is
+#      what lets a silently 4x-slower job go unnoticed. It is not a way out of
+#      the derived requirement: a run that has an own target dir to wipe is a
+#      run whose seed reuses it and clones nothing, so that requirement is
+#      zero. Falling short of a NON-ZERO derived requirement fails the job
+#      here, naming the shortfall and what was kept instead of it — the seed
+#      would otherwise fail seconds later against a half-unshared staging
+#      tree, which is the failure this pass exists to pre-empt.
 #
 # Reactive-only, with no hard cap on cache size: a workspace's natural working
 # set is what it is, and bounding the footprint preemptively means wiping
@@ -75,16 +105,34 @@
 # than reimplementing a lookalike is what makes the classification sound; any
 # drift between two spellings would silently misclassify every directory.
 #
+# The merged half reads the TIP SHA out of that same `ls-remote` output and
+# asks `git merge-base --is-ancestor` against each protected branch's tip,
+# using the objects in this job's own checkout. Ancestry is only decidable
+# where the objects are there to decide it, so the answer "I cannot tell"
+# exists and is distinct from "not merged" everywhere it can arise:
+#
+#   - a shallow checkout makes a MISSING object prove nothing, so the whole
+#     signal is withheld rather than read as "no branch is merged";
+#   - a protected tip that is not in the checkout is not used as an anchor;
+#   - a branch tip that is not in the checkout is kept, loudly.
+#
+# A squash or rebase merge leaves no ancestor relationship at all, so its
+# branch reads as live here. That is a missed reclamation, not a wrong one,
+# and the deleted-branch signal still covers it once the branch is removed.
+#
 # Only ever globs inside <cache-root>. Another project's volume is a different
 # Docker named volume and is not mounted in this container at all, so "stays
 # scoped to this repo's cache" holds structurally, not by convention.
 set -euo pipefail
 . "$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)/cache-lib.sh"
 
-ROOT="${1:?usage: prune-cache.sh <cache-root> <own-target-dir> <protected-branches> <min-free-percent>}"
+ROOT="${1:?usage: prune-cache.sh <cache-root> <own-target-dir> <protected-branches> <min-free-percent> [own-key] [base-key] [fallback-dir]}"
 OWN_DIR="${2:?}"
 PROTECTED_REFS="${3:-}"
-MIN_FREE_PCT="${4:-10}"
+MIN_FREE_PCT="${4:-0}"
+OWN_KEY="${5:-}"
+BASE_KEY="${6:-}"
+FALLBACK="${7:-}"
 # Mirrored by daniel/gitdan's ci-cache-reclaim.sh, whose copy must be >= this
 # one — raising this without raising theirs first lets that script treat a lock
 # this side still honours as abandoned. Same direction and same reasoning as
@@ -102,6 +150,34 @@ STALE_LOCK_SECONDS="${STALE_LOCK_SECONDS:-7200}"
 # the volume is under pressure.
 EVICTION_ASIDE_SETTLE_SECONDS="${EVICTION_ASIDE_SETTLE_SECONDS:-60}"
 
+# What the seed step is about to do, resolved through the same function that
+# step resolves it with (cache-lib.sh's seed_source_candidates). Empty when it
+# will clone nothing at all: its own target dir already exists and it reuses
+# it, or no source exists and it starts cold. Either way the derived
+# requirement is zero, because nothing is about to be copied.
+#
+# gb() is for reporting only. Every comparison below is in KB, because
+# `read_df` reports KB and rounding a threshold to a tenth of a GB either
+# passes a run that cannot fit or evicts a cache the run did not need.
+gb() { awk -v k="${1:-0}" 'BEGIN { printf "%.1f", k / 1048576 }'; }
+
+SEED_SRC=""
+CLONE_KB=0
+if [ -n "$OWN_KEY" ]; then
+  SEED_SRC=$(seed_clone_source "$ROOT" "$OWN_KEY" "$BASE_KEY" "$FALLBACK")
+  if [ -n "$SEED_SRC" ]; then
+    # A full walk of the source, and the reason this pass moved ahead of the
+    # seed rather than staying where it was: the number is only useful before
+    # the clone it describes.
+    CLONE_KB=$(clone_headroom_kb "$SEED_SRC")
+    echo "clone requirement: seeding from $(basename "$SEED_SRC") needs $(gb "$CLONE_KB") GB free (measured from its mutable set)"
+  else
+    echo "clone requirement: none — this run reuses its own cache or starts cold, so nothing will be copied"
+  fi
+else
+  echo "clone requirement: not derivable — no cache key was passed to this pass; the ${MIN_FREE_PCT}% floor is the only gate"
+fi
+
 declare -A protected_ns=()
 for ref in $PROTECTED_REFS; do
   suffix=$(cache_key "$ref")
@@ -109,14 +185,26 @@ for ref in $PROTECTED_REFS; do
   protected_ns["snapshot-${suffix}"]=1
 done
 
-is_protected() {
+# Prints why <dir> is off limits to every pass, or nothing when it is a
+# candidate. The reason is not decoration: it is what the failure report at
+# the bottom lists against each directory it kept while running out of space.
+#
+# SEED_SRC is the third exclusion and the one this script did not used to need.
+# The prune ran after the seed, so the source had already been cloned and the
+# reader marker over it was gone; running BEFORE the seed puts the directory
+# this run is about to read squarely in the candidate set, and pass 1 would
+# take it the moment its branch merged.
+protected_reason() {
   local dir="$1" name
   name=$(basename "$dir")
-  [ "$dir" = "$OWN_DIR" ] && return 0
-  [ -n "${protected_ns[$name]:-}" ] && return 0
+  [ "$dir" = "$OWN_DIR" ] && { printf 'this run own cache'; return 0; }
+  [ -n "$SEED_SRC" ] && [ "$dir" = "$SEED_SRC" ] && { printf 'the source this run is about to clone'; return 0; }
+  [ -n "${protected_ns[$name]:-}" ] && { printf 'a protected branch cache'; return 0; }
   return 1
 }
 
+is_protected() { protected_reason "$1" >/dev/null; }
+
 # is_locked <dir> [name]
 #
 # `name` is the directory's own name for reporting and for the reader-marker
@@ -277,6 +365,11 @@ done
 
 echo "=== pass 1: liveness (unconditional, not gated on free space) ==="
 declare -A live_ns=()
+# The tip SHA origin reports for the branch each directory name belongs to.
+# Same output, same loop, one field over — reading it from a second `git` call
+# would be reading a different instant.
+declare -A live_tip=()
+declare -A remote_tip_of=()
 LIVENESS_AVAILABLE=0
 LIVENESS_REASON=""
 if [ "${CACHE_LIVENESS:-true}" = "0" ] || [ "${CACHE_LIVENESS:-true}" = "false" ]; then
@@ -289,9 +382,13 @@ elif remote_heads=$(timeout 20 git ls-remote --heads origin 2>&1); then
     case "$line" in *refs/heads/*) ;; *) continue ;; esac
     branch="${line#*refs/heads/}"
     [ -n "$branch" ] || continue
+    sha="${line%%[[:space:]]*}"
     suffix=$(cache_key "$branch")
     live_ns["target-${suffix}"]=1
     live_ns["snapshot-${suffix}"]=1
+    live_tip["target-${suffix}"]="$sha"
+    live_tip["snapshot-${suffix}"]="$sha"
+    remote_tip_of["$branch"]="$sha"
     branch_count=$((branch_count + 1))
   done <<< "$remote_heads"
   echo "liveness: ${branch_count} live branches on origin"
@@ -299,6 +396,77 @@ else
   LIVENESS_REASON="git ls-remote --heads origin failed or timed out"
 fi
 
+# The merged half of pass 1, and whether this checkout can answer it at all.
+# Every branch of this decision that ends in "no" ends in the signal being
+# WITHHELD, never in a directory being classified dead by default.
+MERGED_AVAILABLE=0
+MERGED_REASON=""
+PROTECTED_TIPS=()
+declare -A merged_verdict=()
+declare -A merged_into=()
+MERGED_INTO=""
+if [ "$LIVENESS_AVAILABLE" = "1" ]; then
+  if ! git rev-parse --git-dir >/dev/null 2>&1; then
+    MERGED_REASON="not inside a git checkout"
+  elif [ "$(git rev-parse --is-shallow-repository 2>/dev/null || echo unknown)" != "false" ]; then
+    # In a shallow clone an absent commit is the normal case, so `--is-ancestor`
+    # answers about the graph that was fetched rather than the one that exists.
+    MERGED_REASON="the checkout is shallow, so a commit missing from it says nothing about ancestry"
+  else
+    for ref in $PROTECTED_REFS; do
+      tip="${remote_tip_of[$ref]:-}"
+      [ -n "$tip" ] || continue
+      git cat-file -e "${tip}^{commit}" 2>/dev/null || continue
+      PROTECTED_TIPS+=("${ref}:${tip}")
+    done
+    if [ "${#PROTECTED_TIPS[@]}" -gt 0 ]; then
+      MERGED_AVAILABLE=1
+      echo "liveness: merged-branch detection anchored on ${PROTECTED_TIPS[*]%%:*}"
+    else
+      MERGED_REASON="none of the protected branch tips (${PROTECTED_REFS:-none configured}) is present in this checkout"
+    fi
+  fi
+  [ "$MERGED_AVAILABLE" = "1" ] || \
+    echo "::warning::liveness: ${MERGED_REASON} — merged-but-undeleted branches keep their caches this run"
+fi
+
+# is_merged_dead <dir-name>
+#
+# True when the branch this directory belongs to is still on origin but every
+# commit it holds is already on a protected branch — a merged PR whose branch
+# the forge did not delete, which is the case zemyna hits on every merge and
+# which the deleted-branch signal above can never see.
+#
+# Memoised per tip because target-<key> and snapshot-<key> share one branch,
+# and because the "cannot tell" warning belongs to the branch rather than to
+# each of its directories.
+is_merged_dead() {
+  local name="$1" tip="${live_tip[$1]:-}" entry ref psha
+  MERGED_INTO=""
+  [ "$MERGED_AVAILABLE" = "1" ] || return 1
+  [ -n "$tip" ] || return 1
+  case "${merged_verdict[$tip]:-}" in
+    dead) MERGED_INTO="${merged_into[$tip]}"; return 0 ;;
+    live|unknown) return 1 ;;
+  esac
+  if ! git cat-file -e "${tip}^{commit}" 2>/dev/null; then
+    merged_verdict["$tip"]=unknown
+    echo "::warning::prune: ${name}: its branch tip ${tip} is not in this checkout — cannot tell merged from live, keeping it"
+    return 1
+  fi
+  for entry in "${PROTECTED_TIPS[@]}"; do
+    ref="${entry%%:*}"; psha="${entry#*:}"
+    if git merge-base --is-ancestor "$tip" "$psha" 2>/dev/null; then
+      merged_verdict["$tip"]=dead
+      merged_into["$tip"]="$ref"
+      MERGED_INTO="$ref"
+      return 0
+    fi
+  done
+  merged_verdict["$tip"]=live
+  return 1
+}
+
 if [ "$LIVENESS_AVAILABLE" = "1" ]; then
   pruned_any=0
   # Tracked separately so the line below cannot contradict the decline lines
@@ -309,12 +477,20 @@ if [ "$LIVENESS_AVAILABLE" = "1" ]; then
     [ -d "$dir" ] || continue
     name=$(basename "$dir")
     is_protected "$dir" && continue
-    [ -n "${live_ns[$name]:-}" ] && continue
+    if [ -z "${live_ns[$name]:-}" ]; then
+      why="no matching branch on origin"
+      why_summary="branch no longer exists on origin"
+    elif is_merged_dead "$name"; then
+      why="merged into ${MERGED_INTO}, whose tip already contains its every commit"
+      why_summary="merged into \`${MERGED_INTO}\`"
+    else
+      continue
+    fi
     is_locked "$dir" && { spared_any=1; continue; }
     dir_gb=$(usage_gb "$dir")
     evict_dir "$dir" || { spared_any=1; continue; }
-    echo "::warning::pruned dead-branch cache ${name} (${dir_gb} GB) — no matching branch on origin"
-    summary_line "- pruned dead-branch cache \`${name}\` (${dir_gb} GB) — branch no longer exists on origin"
+    echo "::warning::pruned dead-branch cache ${name} (${dir_gb} GB) — ${why}"
+    summary_line "- pruned dead-branch cache \`${name}\` (${dir_gb} GB) — ${why_summary}"
     pruned_any=1
   done
   if [ "$pruned_any" = "0" ]; then
@@ -329,16 +505,45 @@ else
 fi
 
 echo
-echo "=== pass 2/3: disk pressure (threshold: free < ${MIN_FREE_PCT}%) ==="
+echo "=== pass 2/3: disk pressure ==="
 read -r TOTAL_KB FREE_KB <<< "$(read_df "$ROOT")"
-THRESHOLD_KB=$(( TOTAL_KB * MIN_FREE_PCT / 100 ))
+PCT_KB=$(( TOTAL_KB * MIN_FREE_PCT / 100 ))
 
-if [ "$FREE_KB" -ge "$THRESHOLD_KB" ]; then
-  echo "cache: $(basename "$OWN_DIR") $(usage_gb "$OWN_DIR") GB | $(report_df host "$FREE_KB" "$TOTAL_KB")"
+# The two floors, and which of them governs. They are kept apart all the way
+# down rather than collapsed here, because falling short of them means
+# different things: the derived one predicts that the very next step cannot
+# finish, and the percentage one is a hygiene target for the volume.
+REQUIRED_KB="$CLONE_KB"
+GOVERNS="the clone this run is about to make"
+if [ "$PCT_KB" -gt "$REQUIRED_KB" ]; then
+  REQUIRED_KB="$PCT_KB"
+  GOVERNS="the ${MIN_FREE_PCT}% floor"
+fi
+echo "required: $(gb "$REQUIRED_KB") GB free — ${GOVERNS} (clone $(gb "$CLONE_KB") GB, floor $(gb "$PCT_KB") GB)"
+
+own_report() {
+  if [ -d "$OWN_DIR" ]; then
+    echo "cache: $(basename "$OWN_DIR") $(usage_gb "$OWN_DIR") GB | $(report_df host "$1" "$2")"
+  else
+    # Ordinary now that this pass runs ahead of the seed: on a branch's first
+    # run of the day the directory does not exist yet, and reporting 0.0 GB
+    # for it would read as an emptied cache.
+    echo "cache: $(basename "$OWN_DIR") not seeded yet | $(report_df host "$1" "$2")"
+  fi
+}
+
+if [ "$FREE_KB" -ge "$REQUIRED_KB" ]; then
+  own_report "$FREE_KB" "$TOTAL_KB"
   exit 0
 fi
 
-echo "::warning::$(report_df disk "$FREE_KB" "$TOTAL_KB") < ${MIN_FREE_PCT}% threshold"
+echo "::warning::$(report_df disk "$FREE_KB" "$TOTAL_KB") < the $(gb "$REQUIRED_KB") GB this run requires"
+
+# What survived the pass, and why, in the order the pass considered them. Read
+# only by the failure report at the bottom: a run that cannot fit its clone is
+# a run whose log has to answer "then what is all that space?" without anyone
+# having to reconstruct the pass by hand.
+KEPT=()
 
 # A plain loop over a pre-materialised, pre-sorted list rather than a live
 # `find | while` pipeline, so `rm -rf` inside the loop cannot make a running
@@ -346,21 +551,56 @@ echo "::warning::$(report_df disk "$FREE_KB" "$TOTAL_KB") < ${MIN_FREE_PCT}% thr
 mapfile -t LRU < <(list_by_lru)
 for dir in "${LRU[@]}"; do
   [ -d "$dir" ] || continue
-  is_protected "$dir" && continue
+  if reason=$(protected_reason "$dir"); then
+    KEPT+=("$(basename "$dir") — ${reason}")
+    continue
+  fi
   read -r TOTAL_KB FREE_KB <<< "$(read_df "$ROOT")"
-  [ "$FREE_KB" -ge "$THRESHOLD_KB" ] && break
-  is_locked "$dir" && continue
+  [ "$FREE_KB" -ge "$REQUIRED_KB" ] && break
+  if is_locked "$dir"; then
+    KEPT+=("$(basename "$dir") — held open by a running job")
+    continue
+  fi
   dir_gb=$(usage_gb "$dir")
-  evict_dir "$dir" || continue
+  if ! evict_dir "$dir"; then
+    KEPT+=("$(basename "$dir") — claimed by a job while its eviction was in flight")
+    continue
+  fi
   echo "::warning::evicted $(basename "$dir") (${dir_gb} GB, LRU under disk pressure)"
   summary_line "- evicted \`$(basename "$dir")\` (${dir_gb} GB, LRU under disk pressure)"
 done
 
 read -r TOTAL_KB FREE_KB <<< "$(read_df "$ROOT")"
-if [ "$FREE_KB" -lt "$THRESHOLD_KB" ]; then
+
+# Falling short of the DERIVED requirement is a failure, not a warning. The
+# seed step is next, it will clone that source, and it will run out of disk
+# part-way through unsharing the clone's mutable paths — reported against a
+# staging path, with nothing in the message about which cache was holding the
+# space. Failing here says that instead.
+#
+# There is nothing to self-clear on this path and it is not skipped in error:
+# a non-zero requirement means the seed is about to CLONE, which means this
+# run has no own target dir to wipe (seed_clone_source returns nothing when it
+# does), so pass 3 has no candidate. See the header.
+if [ "$FREE_KB" -lt "$CLONE_KB" ]; then
+  echo "::error::prune: $(gb "$FREE_KB") GB free after evicting every eligible cache, but seeding from $(basename "$SEED_SRC") needs $(gb "$CLONE_KB") GB — short by $(gb "$(( CLONE_KB - FREE_KB ))") GB"
+  summary_line "- **out of disk**: seeding from \`$(basename "$SEED_SRC")\` needs $(gb "$CLONE_KB") GB, $(gb "$FREE_KB") GB free"
+  echo "prune: kept, and why:"
+  for entry in ${KEPT[@]+"${KEPT[@]}"}; do echo "  ${entry}"; done
+  [ "${#KEPT[@]}" -gt 0 ] || echo "  (nothing — the volume holds no cache directories at all)"
+  exit 1
+fi
+
+if [ "$FREE_KB" -lt "$PCT_KB" ]; then
   OWN_GB=$(usage_gb "$OWN_DIR")
   echo "::warning::still under threshold after evicting every eligible sibling; clearing own $(basename "$OWN_DIR") (was ${OWN_GB} GB) — this run pays a cold rebuild"
   summary_line "- **self-clear**: \`$(basename "$OWN_DIR")\` (was ${OWN_GB} GB) wiped — this run pays a cold rebuild"
+  # Recreated empty rather than left absent, and that is what keeps this path
+  # out of the requirement above: the seed reuses an own target dir that
+  # exists, whatever is in it, so a self-cleared run clones nothing and needs
+  # no headroom. Leaving it absent would send that run to the base snapshot
+  # instead, needing a clone this pass has just spent its last eligible bytes
+  # not sizing for.
   rm -rf "$OWN_DIR"
   mkdir -p "$OWN_DIR"
 else

From 24f87a6b98ff8379a5fb69aa4951c9844fa529f6 Mon Sep 17 00:00:00 2001
From: Claude <claude@gitdan.com>
Date: Mon, 7 Sep 2026 00:21:12 -0500
Subject: [PATCH 3/4] docs(cache): describe both liveness signals and the
 derived requirement
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The eviction section described one way for a branch to be dead and a
percentage threshold that no longer exists as a gate. Rewritten around
what the pass actually does: two dead-branch signals with the limits of
the ancestry one stated, a requirement measured off the clone's mutable
set, and a failure that names its shortfall.

The 10% figure is deleted rather than corrected — it was the default of
a gate, and min-free-percent is now an additional floor defaulting to 0,
so there is no percentage left to state. The inputs table, the prune and
liveness rows, the selftest coverage row and the fetch-depth comment in
the usage example all say what they now mean; fetch-depth: 0 has a
second reason to be required, since a shallow checkout cannot answer
ancestry.

Refs #20.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JXMQCJ5Eg5f9G9cfYzyh4Z
---
 README.md | 81 ++++++++++++++++++++++++++++++++++++++++++++-----------
 1 file changed, 65 insertions(+), 16 deletions(-)

diff --git a/README.md b/README.md
index 273fea4..4b55110 100644
--- a/README.md
+++ b/README.md
@@ -111,10 +111,14 @@ jobs:
     steps:
       - uses: actions/checkout@v4
         with:
-          # REQUIRED. The mtime restore walks every commit that ever touched a
-          # tracked file; a depth-1 checkout makes every file resolve to the
-          # tip commit and the cache stops working. The action fails loudly
-          # rather than silently degrading if this is missing.
+          # REQUIRED, for two things. The mtime restore walks every commit
+          # that ever touched a tracked file; a depth-1 checkout makes every
+          # file resolve to the tip commit and the cache stops working, and
+          # the action fails loudly rather than silently degrading. The prune
+          # also decides whether a branch has been merged by asking this
+          # checkout for ancestry, which a shallow one cannot answer — there
+          # it withholds that half of the pass and says so, so merged-but-
+          # undeleted branches keep their caches.
           fetch-depth: 0
 
       - uses: https://gitdan.com/daniel/gitdan-actions/cargo-cache@v1
@@ -263,14 +267,59 @@ not collapsed into one:
   disk, normally as the cache restored under its own name and otherwise as one
   set aside for a later pass to reclaim.
 
-**Eviction** runs three passes: caches for branches that no longer exist on
-origin are removed unconditionally; then, only if free space is under the
-threshold, live caches are evicted oldest-first; then, as a last resort, this
-run's own cache. Protected refs and any cache held open by a running job are
-never candidates. Within the pressure pass, `target-*` directories are evicted
-before `snapshot-*` ones — the reverse of the obvious order, because a
-snapshot is hardlinked to everything cloned from it, so removing one frees
-almost no real bytes while costing every future PR its warm start.
+**Eviction** runs three passes, ahead of the seed so that what it frees is
+available to the clone that follows. Caches for branches that are DEAD are
+removed unconditionally; then, only if free space is under the requirement,
+live caches are evicted oldest-first; then, as a last resort, this run's own
+cache. Protected refs, the source this run is about to clone, and any cache
+held open by a running job are never candidates. Within the pressure pass,
+`target-*` directories are evicted before `snapshot-*` ones — the reverse of
+the obvious order, because a snapshot is hardlinked to everything cloned from
+it, so removing one frees almost no real bytes while costing every future PR
+its warm start.
+
+**A branch is dead in two ways, and the second is the one that reclaims
+anything here.** Gitea keeps a PR's branch after the merge unless the repo
+opts into delete-on-merge, so `git ls-remote` reports merged branches forever
+and "gone from origin" fires for almost nothing. The second signal is
+ancestry: a branch still on origin whose tip is an ancestor of a protected
+branch's tip holds no commit that branch does not, so its cache will never be
+read again and goes in the same pass. On zemyna's volume that is the
+difference between reclaiming nothing and reclaiming a 40 GB directory per
+merged PR (gitdan-actions#20).
+
+Ancestry is answered from the commits in the job's own checkout, so it is only
+answered where they are there to answer it — and "cannot tell" is never folded
+into "dead", at either granularity. A shallow checkout withholds the signal
+entirely, because a missing object is its normal case rather than evidence; a
+single branch whose tip is not in the checkout is kept, with a warning naming
+it. A squash or rebase merge leaves no ancestry at all, so its branch reads as
+live until it is deleted. All three are missed reclamations, which cost disk;
+the alternative direction costs a branch its cache while it is still being
+built on.
+
+**How much free space is enough is measured, not chosen.** What the seed is
+about to do is hardlink-clone a snapshot and then real-copy that clone's
+*mutable set* — the dep-info, build-script metadata and linked outputs that a
+build would otherwise write through a shared inode. The rest stays hardlinked
+and costs nothing. So the requirement is derived per run, from that set,
+measured off the very snapshot the seed will read using the same enumeration
+`unshare_mutable_paths` copies from; the pass evicts oldest-first until it is
+met and then stops. A percentage of the volume cannot express this: how much a
+clone needs is a property of the snapshot, and a threshold sized for a
+different failure is one that never fires before the seed refuses. `cp -al`
+still materialises every directory for real, and the unshare stages each
+subtree through a sibling copy, so the measurement carries a margin —
+`CACHE_CLONE_HEADROOM_PERCENT` and `CACHE_CLONE_HEADROOM_FLOOR_KB`, both
+hand-written defaults, both erring toward asking for more.
+
+A run that cannot reach the derived requirement after evicting everything
+eligible **fails, naming the shortfall and every directory it kept instead**.
+The seed would otherwise fail seconds later, reporting a staging path and
+nothing about which cache was holding the space — which is the failure this
+pass now pre-empts. `min-free-percent` is an additional floor on top and
+nothing more: it defaults to `0`, it only ever raises the requirement, and
+falling short of it is still a warning and a self-clear rather than a failure.
 
 **File mtimes.** `actions/checkout` stamps every file with "now", which makes
 every crate look changed to Cargo's mtime-based freshness check — a persistent
@@ -412,10 +461,10 @@ directory; the line just doesn't say which.
 | `cache-root` | `/cache` | mount point of the persistent volume inside the job container |
 | `cache-lineage` | *(empty)* | one directory level under `cache-root`, for a second job building the same ref for a different target or profile — see [Multiple jobs in one workflow](#multiple-jobs-in-one-workflow) |
 | `protected-branches` | `dev main` | refs that publish snapshots and are never evicted |
-| `min-free-percent` | `10` | prune when free space drops below this |
+| `min-free-percent` | `0` | an ADDITIONAL free-space floor, as a percentage of the volume. The gate is derived per run from what the seed is about to clone; this only ever raises it |
 | `restore-mtimes` | `true` | restore tracked-file mtimes from git history |
-| `prune` | `true` | run the eviction pass |
-| `liveness-prune` | `true` | within eviction, remove caches for branches gone from origin |
+| `prune` | `true` | run the eviction pass — before the seed, so what it frees is available to the clone |
+| `liveness-prune` | `true` | within eviction, remove caches for branches that are dead: gone from origin, or merged into a protected branch |
 | `own-ref` | *(auto)* | override; defaults to `github.head_ref`, else `github.ref_name` |
 | `base-ref` | *(auto)* | override; defaults to `github.base_ref` (empty on push) |
 | `seed-fallback-dir` | *(empty)* | absolute path to seed from when no snapshot exists — for migrating off an existing flat cache |
@@ -652,7 +701,7 @@ change here reaches all of them at once. That is what the gate is for.
 | `hardlink-clone-selftest.sh` | that a build in a clone cannot mutate its source — with a control proving a raw `cp -al` does. Needs a real compiler, **and a nightly that actually resolves freshness by content for its last scenario**: the source's-next-build check reasons about content rather than mtime, so under mtime freshness it would assert a bug. Whether the toolchain does is settled by experiment on a throwaway crate, not by asking it — accepting `-Z checksum-freshness` stopped implying it on 2026-08-22, when cargo PR #17382 demoted the flag to a gate and gave `build.fingerprint` (default `mtime`) the choice; the suite now exports both and 1.100.0-nightly measures ACTIVE again. The experiment reports **three** outcomes, not two: active, measured-inactive, and *not measured*. Its answer codes are `0` and `3`, deliberately clear of every status bash generates for its own errors — so nothing that goes wrong inside the probe, including an expansion failure no guard can catch, can be read as an answer. The scenario is skipped for the last two alike, but a failure to measure is never reported as a measurement. The control also reports which mutation families the running Cargo exhibits — a note, not an assertion, since that set moves upstream. It also pins the SELECTION itself against both of Cargo's build-dir layouts, from two file-only fixtures that need no compiler — so the layout the installed Cargo does not happen to write is still covered — and asserts the partition in both directions: every file Cargo rewrites in place is privately owned, and every `.rlib`/`.rmeta` still shares its inode. The second half is what the old suite never checked beyond `shared > 0`, and it is what a layout change silently inverts. |
 | `seed-target-dir-selftest.sh` | seed-source preference, lock-file stripping, two jobs racing on one cache key, **and one scenario per check a hardlink clone is validated against**: a source rotated wholesale, a subtree silently lost from the walk, a copy that reports failure over a tree both other checks read as whole, and a source identity that resolved at neither end — plus a staging tree that could not be privately owned being discarded rather than published, and the publisher's log showing it waited on the consumer's own reader-lock marker before reclaiming a rotated snapshot |
 | `publish-snapshot-selftest.sh` | the atomic swap, that a live consumer survives a republish, and the publisher's side of the rotation race: deferred reclamation under a live reader, and its sweep once the reader is gone |
-| `prune-cache-selftest.sh` | liveness, protection, locking, eviction order, self-clear, **and that a cache a job claims *inside* the check-to-unlink window survives it** — against a real scratch `origin` |
+| `prune-cache-selftest.sh` | liveness in both its forms — a branch deleted from origin, and one still on it whose tip is already merged — plus protection, locking, eviction order, self-clear, **that a cache a job claims *inside* the check-to-unlink window survives it**, and that a requirement derived from the clone's mutable set evicts exactly enough and then fails rather than under-delivering. Against a real scratch `origin`, including a genuinely shallow clone of it and a `df` that answers from the fixture's own size, since a fixed one cannot show a pass stopping |
 | `restore-mtimes-selftest.sh` | the merge hazard and the watermark that closes it, including the two-jobs-one-namespace case. Needs a real compiler. |
 
 Every suite runs the actual script, not a reimplementation of its logic, and

From 38a63879365ed3aaa152750b7d11a4b4638f5f05 Mon Sep 17 00:00:00 2001
From: Claude <claude@gitdan.com>
Date: Mon, 7 Sep 2026 00:27:11 -0500
Subject: [PATCH 4/4] docs(cache): recommend delete-on-merge, and say what
 ancestry still covers
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

`daniel/zemyna` enabled `default_delete_branch_after_merge` after this
branch was written, so the deleted-branch signal will fire there on
future merges. That makes the setting worth recommending — it is the
cheapest case for this scheme, decidable from `ls-remote` with no
checkout, no objects and no walk — and it does not make the ancestry
signal redundant.

Three things it leaves behind, now enumerated in the README's eviction
section rather than implied: every branch merged before the setting was
turned on, of which zemyna carried 48 and which nothing retroactively
deletes; every merge whose deletion the forge declines or is never asked
to make, since it is best-effort and silent and an API merge without the
flag never asks; and every repo that has not enabled it, which is the
default.

Two present-tense claims about one repo's configuration are reworded
into the conditions they were standing in for, in prune-cache.sh's
header and beside is_merged_dead, plus the two in the selftest that
asserted the forge keeps branches rather than describing the fixture.
No behaviour change; the suite is green and unchanged at 61 assertions.

Refs #20.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JXMQCJ5Eg5f9G9cfYzyh4Z
---
 README.md                       | 28 +++++++++++++++++++++-------
 scripts/prune-cache-selftest.sh | 10 +++++-----
 scripts/prune-cache.sh          | 12 ++++++++----
 3 files changed, 34 insertions(+), 16 deletions(-)

diff --git a/README.md b/README.md
index 4b55110..dff0b21 100644
--- a/README.md
+++ b/README.md
@@ -278,15 +278,29 @@ the obvious order, because a snapshot is hardlinked to everything cloned from
 it, so removing one frees almost no real bytes while costing every future PR
 its warm start.
 
-**A branch is dead in two ways, and the second is the one that reclaims
-anything here.** Gitea keeps a PR's branch after the merge unless the repo
-opts into delete-on-merge, so `git ls-remote` reports merged branches forever
-and "gone from origin" fires for almost nothing. The second signal is
+**A branch is dead in two ways, and neither signal makes the other
+redundant.** The first is that the branch is gone from origin. The second is
 ancestry: a branch still on origin whose tip is an ancestor of a protected
 branch's tip holds no commit that branch does not, so its cache will never be
-read again and goes in the same pass. On zemyna's volume that is the
-difference between reclaiming nothing and reclaiming a 40 GB directory per
-merged PR (gitdan-actions#20).
+read again and goes in the same pass.
+
+**Turn delete-on-merge on** (`default_delete_branch_after_merge`, per repo) —
+it is the setting this scheme is cheapest under, because a deleted branch is
+decidable from `ls-remote` alone, with no checkout, no objects and no walk.
+The ancestry signal is what covers the rest, and the rest is not a corner:
+
+- **Every branch merged before the setting was turned on.** They stay on
+  origin forever; nothing retroactively deletes them. zemyna carried 48 of
+  them at the time the setting was enabled, and ancestry is the only thing
+  that reclaims a cache dir belonging to any of them.
+- **Every merge the deletion declines or fails.** Gitea's delete is
+  best-effort and silent: it declines for a protected branch and for one
+  another open PR still uses, and an API merge that omits the flag — which
+  `tea pulls merge` does — simply never asks.
+- **Repos that have not enabled it**, which is the default.
+
+That is the difference between reclaiming nothing and reclaiming a 40 GB
+directory per merged PR on a full volume (issue 20).
 
 Ancestry is answered from the commits in the job's own checkout, so it is only
 answered where they are there to answer it — and "cannot tell" is never folded
diff --git a/scripts/prune-cache-selftest.sh b/scripts/prune-cache-selftest.sh
index ff5cca8..e83e239 100755
--- a/scripts/prune-cache-selftest.sh
+++ b/scripts/prune-cache-selftest.sh
@@ -46,10 +46,10 @@
 #      empty a tree its owner may still restore under a live cache name. Its
 #      fixture is an OLD directory renamed a moment ago — production's shape,
 #      and what lets it tell the two timestamps apart.
-#  15. A MERGED-BUT-UNDELETED BRANCH IS DEAD TOO. This forge keeps branches
-#      after merge, so `ls-remote` reports them forever and scenario 1's
-#      signal never fires for them — which is how three 40 GB caches sat on a
-#      full volume until somebody removed them by hand (gitdan-actions#20). A
+#  15. A MERGED-BUT-UNDELETED BRANCH IS DEAD TOO. A branch the forge did not
+#      delete at merge stays on `ls-remote` forever, so scenario 1's signal
+#      never fires for it — which is how three 40 GB caches sat on a full
+#      volume until somebody removed them by hand (gitdan-actions#20). A
 #      branch whose tip is an ancestor of a protected branch's tip is pruned
 #      like a deleted one; an unmerged branch beside it is not.
 #  16. AND "CANNOT TELL" IS STILL NOT DEATH, at both granularities: a branch
@@ -314,7 +314,7 @@ assert_log "may still be evicting it" "the deferral gives its actual reason"
 
 echo
 echo "=== 15: a merged-but-undeleted branch is dead too ==="
-# This forge keeps a PR's branch after the merge, so `ls-remote` reports it
+# A branch the forge did not delete at merge, which `ls-remote` then reports
 # forever. Built the way that happens: a branch merged into dev with a merge
 # commit, still pushed, beside one branched at the same point and NOT merged.
 git="git -c user.email=t@t -c user.name=t -c commit.gpgsign=false"
diff --git a/scripts/prune-cache.sh b/scripts/prune-cache.sh
index 7e1827c..bf20f4e 100755
--- a/scripts/prune-cache.sh
+++ b/scripts/prune-cache.sh
@@ -28,8 +28,9 @@
 #      removed UNCONDITIONALLY, not gated on free space. A directory for a
 #      branch nothing will build again is pure loss; waiting for disk pressure
 #      to notice means paying for it until then. Two signals make a branch
-#      dead, and the second exists because the first alone is inert on a forge
-#      that keeps branches after merge (gitdan-actions#20):
+#      dead, and the second exists because the first alone is inert wherever
+#      a merged branch stays on origin — the default, and still the outcome
+#      whenever delete-on-merge declines or is not asked (gitdan-actions#20):
 #
 #        DELETED — the branch is no longer on origin at all.
 #        MERGED  — the branch is still on origin, but its tip is an ancestor
@@ -434,8 +435,11 @@ fi
 #
 # True when the branch this directory belongs to is still on origin but every
 # commit it holds is already on a protected branch — a merged PR whose branch
-# the forge did not delete, which is the case zemyna hits on every merge and
-# which the deleted-branch signal above can never see.
+# the forge did not delete, which the deleted-branch signal above can never
+# see. Enabling delete-on-merge narrows this to the branches merged before it
+# was enabled, the ones its deletion declines (protected, or used by another
+# open PR), and the merges that never ask (an API merge without the flag);
+# see README's eviction section.
 #
 # Memoised per tip because target-<key> and snapshot-<key> share one branch,
 # and because the "cannot tell" warning belongs to the branch rather than to