fix(wake): #946 digest ack watermark clamped at quarantined seqs — disclose AND clamp
ci/woodpecker/pr/ci Pipeline was successful

The embedded ack suggestion covered dead-lettered entries: quarantine is a
render-time filter (0.6.14), so `ack.sh consumed --upto <observed_seq>`
stepped the cursor past quarantined seqs and _record_last_consumed wrote
consumed-hash witness rows for deliveries that never happened (live: mos-dt
seq 68 buried under five successive digests; Finding A: a false
9d0f639f…@63 witness row).

Fix, per the ruled disposition on #946:

- digest.sh: rendered digest gains a QUARANTINED section (seq + class +
  HELD only — ids and locator values stay withheld, preserving the
  exclusion property); embedded ack clamped to
  min(observed_seq, min quarantined seq - 1) with a loud
  "# ACK CLAMPED (#946)" note; render --from-store syncs the store-owned
  quarantined.set via `store.sh quarantine-sync` (full replace — a gate
  fix self-heals stale quarantine); --from-file/--stdin never touch the
  set.

- store.sh: consume REFUSES to cross an unconsumed quarantined seq;
  `--force-past-quarantine` is the ONLY way past, loud per-seq on stderr,
  and even the forced path never writes a consumed-hash witness for a
  quarantined seq; crossed seqs are pruned from the set after the cursor
  moves. New `quarantine-sync` (stdin seqs, full replace, invalid input is
  a loud no-op) and `quarantine-audit [--repair]` (sweeps consumed-hashes
  for rows provably contradicted by the dead-letter ledger; report exits
  1; --repair removes only provably-false rows; the ledger is history and
  never modified; rows whose dead-letter evidence was pruned are
  unprovable and untouched).

- ack.sh: plumbs --force-past-quarantine through to store consume; forced-
  path loudness is re-emitted on stderr while `CONSUMED <n>` stays clean.

Tests: test-wake-store-ack.sh T13-T16 (34 assertions RED at base),
test-wake-digest-quarantine.sh Q12-Q16 (10 RED at base; Q13/Q16
green-by-design controls). All nine wake suites green. version=0.6.15.

Closes #946

Written-by: pepper (sb-it-1-dt)
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01NsKce8iZuSuRnu3gVMCBKB
This commit is contained in:
Jason Woltje
2026-07-30 10:10:21 -05:00
co-authored by Claude Fable 5
parent a6b5f6a01a
commit e11bc66223
6 changed files with 631 additions and 13 deletions
+208 -6
View File
@@ -77,11 +77,42 @@ Commands:
the actual paste is out of scope.
If --require-idle-cmd is given and it
exits non-zero, emit nothing (not idle).
consume --upto N Advance consumed_seq over the contiguous
consume --upto N [--force-past-quarantine]
Advance consumed_seq over the contiguous
gapless prefix <=N; drop consumed
entries. Rejects a gap (cannot ack N
while N-1 is unconsumed). Cumulative &
idempotent.
idempotent. REFUSES to advance past a
QUARANTINED seq (#946): a quarantined
entry was dead-lettered at render and
never delivered in any digest, so an
ordinary ack may not record it consumed.
--force-past-quarantine is the ONLY way
past — loud per stepped-over seq, and
even then NO consumed-hash witness is
recorded for the quarantined entry.
quarantine-sync REPLACE the store-owned quarantined.set
with the observed_seqs read from stdin
(one per line; empty input CLEARS).
Called by digest.sh after each
authoritative store render — the set is
re-DERIVED per render, never accumulated,
so a fixed locator gate self-heals the
consume clamp (#944 recovery).
quarantine-audit [--repair] Report consumed-hashes rows that are
PROVABLY FALSE: the row matches a
dead-letter ledger entry on
(kind,id,observed_seq,observed_hash) at
or below consumed_seq — i.e. the recorded
consumption was of a quarantined,
never-delivered entry (#946 Finding A).
Report-only by default (exit 1 when any
found); --repair removes exactly those
rows (atomic, loud). The dead-letter
ledger itself is NEVER modified (it is
history). Rows whose dead-letter evidence
was pruned are NOT provable and are
never touched.
cursors Print observed_seq / consumed_seq / depth.
Environment:
@@ -354,10 +385,20 @@ cmd_drain() {
# (the reconciler re-enumerates the consumed state once — no lost obligation),
# never a failed consume.
_record_last_consumed() {
local upto="$1" existing new_records merged
local upto="$1" existing new_records merged qjson
[ -f "$STATE_DIR/pending.jsonl" ] || return 0
new_records="$(jq -c --argjson upto "$upto" '
select((.observed_seq // -1) <= $upto)
# #946: seqs in quarantined.set are EXCLUDED from the record — the trust
# boundary above says "existence implies durably enqueued+CONSUMED", but a
# quarantined entry was dead-lettered at render and NEVER delivered, so a row
# for it would witness a delivery that never happened. This holds even on the
# FORCED step-over path: the reconciler re-enumerating the state once is
# safe-but-noisy; a false witness silences it forever.
qjson="$(jq -nR -c '[inputs | select(length > 0) | tonumber? // empty]' "$STATE_DIR/quarantined.set" 2>/dev/null || true)"
[ -n "$qjson" ] || qjson='[]'
new_records="$(jq -c --argjson upto "$upto" --argjson quarantined "$qjson" '
(.observed_seq // -1) as $seq
| select($seq <= $upto)
| select(($quarantined | index($seq)) == null)
| select((.locators.kind // "") != "" and (.locators.id // "") != "" and (.locators.observed_hash // "") != "")
| {kind:.locators.kind, id:.locators.id, observed_hash:.locators.observed_hash, observed_seq:.observed_seq}
' "$STATE_DIR/pending.jsonl" 2>/dev/null || true)"
@@ -372,13 +413,17 @@ _record_last_consumed() {
}
cmd_consume() {
local upto=''
local upto='' force=0
while [ $# -gt 0 ]; do
case "$1" in
--upto)
upto="${2:-}"
shift 2
;;
--force-past-quarantine)
force=1
shift
;;
*)
echo "store.sh consume: unknown option '$1'" >&2
exit 2
@@ -426,6 +471,33 @@ cmd_consume() {
k=$((k + 1))
done
# --- #946 quarantine CLAMP -------------------------------------------------
# A quarantined seq was DEAD-LETTERED at render (no §2.1 hard locator): it was
# never delivered in any digest, so advancing consumed_seq past it would record
# consumption of an entry the consumer has never seen. The ordinary path
# REFUSES; --force-past-quarantine is the ONLY way past, and it is loud per
# stepped-over seq. digest.sh re-derives the set at each authoritative store
# render (quarantine-sync REPLACE), so a fixed locator gate self-heals this
# clamp without operator action.
local qfile="$STATE_DIR/quarantined.set" blocked='' q
if [ -s "$qfile" ]; then
while IFS= read -r q; do
case "$q" in '' | *[!0-9]*) continue ;; esac
if [ "$q" -gt "$consumed" ] && [ "$q" -le "$upto" ]; then
blocked="$blocked $q"
fi
done <"$qfile"
fi
if [ -n "$blocked" ]; then
if [ "$force" -eq 0 ]; then
echo "store.sh consume: REFUSED (#946 quarantine clamp) — quarantined seq(s):$blocked inside (consumed_seq=$consumed, upto=$upto] were dead-lettered at render and NEVER delivered in any digest. Advancing past them would record consumption of entries the consumer has never seen. Disposition them (see $STATE_DIR/dead-letter.jsonl) or step over EXPLICITLY with --force-past-quarantine." >&2
exit 1
fi
for q in $blocked; do
echo "store.sh consume: FORCED PAST QUARANTINE (#946) — stepping consumed_seq over quarantined seq $q (dead-lettered, NEVER delivered). No consumed-hash witness is recorded for it; the obligation stays visible ONLY in $STATE_DIR/dead-letter.jsonl." >&2
done
fi
# #932: record the last-consumed observed_hash per (kind,id) BEFORE the pending
# prefix is dropped (this reads the entries about to be truncated), so the
# reconciler can recognise an already-consumed state as ACCOUNTED instead of
@@ -439,9 +511,137 @@ cmd_consume() {
awk -v c="$upto" 'NF && $1+0 > c' "$STATE_DIR/observed.set" 2>/dev/null |
_atomic_write "$STATE_DIR/observed.set" || true
printf '%s' "$upto" | _atomic_write "$STATE_DIR/consumed_seq"
# #946: on a forced step-over, PRUNE the stepped-over seqs from quarantined.set
# (they are inside the consumed prefix now; a stale entry would re-refuse the
# next consume forever). Mirrors the observed.set prune idiom above.
if [ -n "$blocked" ]; then
awk -v c="$upto" 'NF && $1+0 > c' "$qfile" 2>/dev/null |
_atomic_write "$qfile" || true
fi
echo "$upto"
}
# cmd_quarantine_sync — #946: REPLACE the store-owned quarantined.set with the
# observed_seqs read from stdin (one per line). Called by digest.sh after each
# AUTHORITATIVE full-set render (src=store): the set is re-DERIVED per render,
# never accumulated, so a fixed locator gate self-heals the consume clamp (the
# #944 recovery case — a cumulative-forever set would keep refusing acks on
# entries that now render). Empty input CLEARS the set. Foreign-data renders
# (--from-file/--stdin) never call this. Atomic write; invalid input is refused
# loudly with the set left untouched.
cmd_quarantine_sync() {
[ $# -eq 0 ] || {
echo "store.sh quarantine-sync: takes no options (observed_seqs on stdin, one per line)" >&2
exit 2
}
local seq list=''
while IFS= read -r seq; do
[ -n "$seq" ] || continue
case "$seq" in
*[!0-9]*)
echo "store.sh quarantine-sync: invalid observed_seq '$seq' — one non-negative integer per line; set left untouched" >&2
exit 2
;;
esac
list="$list$seq"$'\n'
done
_wake_init_dir "$STATE_DIR"
if ! { printf '%s' "$list" | grep -v '^[[:space:]]*$' || true; } | sort -n | uniq | _atomic_write "$STATE_DIR/quarantined.set"; then
echo "store.sh quarantine-sync: quarantined.set write FAILED — the consume clamp may be stale for this lane" >&2
exit 1
fi
}
# cmd_quarantine_audit — #946 Finding A: the pre-#946 consume recorded
# consumed-hash rows for QUARANTINED (never-delivered) entries — false
# witnesses that silence the reconciler for keys whose only "consumption" was a
# dead-lettered entry (live: safe by population only where the key re-emitted
# and a later seq won the per-key max_by merge; keys that never recurred keep
# the false row forever).
#
# A row is PROVABLY FALSE iff the dead-letter ledger contains an entry matching
# it on (kind, id, observed_seq, observed_hash) AND row.observed_seq <=
# consumed_seq: the per-key max_by merge means the surviving row's provenance IS
# that quarantined entry (a healed row differs in seq/hash and never matches).
# PROVABILITY BOUND: a row whose dead-letter evidence was pruned/rotated away is
# NOT provable and is never touched — this audit only ever removes what the
# ledger can convict. The dead-letter ledger itself is history and is NEVER
# modified here.
cmd_quarantine_audit() {
local repair=0
while [ $# -gt 0 ]; do
case "$1" in
--repair)
repair=1
shift
;;
*)
echo "store.sh quarantine-audit: unknown option '$1'" >&2
exit 2
;;
esac
done
_need_jq
_wake_init_dir "$STATE_DIR"
local rec="$STATE_DIR/consumed-hashes.jsonl" dlf="$STATE_DIR/dead-letter.jsonl"
if [ ! -s "$rec" ]; then
echo "store.sh quarantine-audit: OK — no consumed-hashes record to audit"
return 0
fi
local dl
dl="$(jq -s -c '[.[] | {kind: (.locators.kind // ""), id: ((.locators.id // "") | tostring), observed_seq: (.observed_seq // -1), observed_hash: (.locators.observed_hash // "")}]' "$dlf" 2>/dev/null || true)"
[ -n "$dl" ] || dl='[]'
local consumed
consumed="$(_wake_read_int "$STATE_DIR/consumed_seq" 0)"
local false_rows
false_rows="$(jq -c --argjson dl "$dl" --argjson consumed "$consumed" '
. as $row
| select(($row.observed_seq // -1) <= $consumed)
| select(($dl | map(select(
.kind == ($row.kind // "")
and .id == (($row.id // "") | tostring)
and .observed_seq == ($row.observed_seq // -1)
and .observed_hash == ($row.observed_hash // "")
)) | length) > 0)
' "$rec" 2>/dev/null || true)"
if [ -z "$false_rows" ]; then
echo "store.sh quarantine-audit: OK — no provably-false consumed-hash rows (rows without surviving dead-letter evidence are not provable and were not judged)"
return 0
fi
local n row
n="$(printf '%s\n' "$false_rows" | grep -c . || true)"
while IFS= read -r row; do
[ -n "$row" ] || continue
if [ "$repair" -eq 1 ]; then
echo "store.sh quarantine-audit: REPAIR — removing FALSE WITNESS row $row (matches a dead-lettered, never-delivered entry at/below consumed_seq=$consumed)" >&2
else
echo "FALSE WITNESS — consumed-hashes row $row matches a dead-lettered, never-delivered entry at/below consumed_seq=$consumed (the recorded consumption never happened)"
fi
done <<<"$false_rows"
if [ "$repair" -eq 0 ]; then
echo "store.sh quarantine-audit: $n provably-false row(s) found — run with --repair to remove exactly these rows"
return 1
fi
local kept
kept="$(jq -c --argjson dl "$dl" --argjson consumed "$consumed" '
. as $row
| select(
(($row.observed_seq // -1) > $consumed)
or (($dl | map(select(
.kind == ($row.kind // "")
and .id == (($row.id // "") | tostring)
and .observed_seq == ($row.observed_seq // -1)
and .observed_hash == ($row.observed_hash // "")
)) | length) == 0)
)
' "$rec" 2>/dev/null || true)"
if ! { printf '%s\n' "$kept" | grep -v '^[[:space:]]*$' || true; } | _atomic_write "$rec"; then
echo "store.sh quarantine-audit: consumed-hashes rewrite FAILED — record left untouched" >&2
exit 1
fi
echo "store.sh quarantine-audit: repaired — removed $n provably-false row(s); dead-letter ledger untouched (history)"
}
cmd_cursors() {
_wake_init_dir "$STATE_DIR"
local observed consumed depth
@@ -463,6 +663,8 @@ main() {
enqueue) cmd_enqueue "$@" ;;
drain) cmd_drain "$@" ;;
consume) cmd_consume "$@" ;;
quarantine-sync) cmd_quarantine_sync "$@" ;;
quarantine-audit) cmd_quarantine_audit "$@" ;;
cursors) cmd_cursors "$@" ;;
-h | --help | help) usage ;;
*)