Task — engineering-spec@1

Uninstall must leave the repo as it found it

doneTASK-IMP-121
module improvement · class product · priority p1 · created 2026-07-19 · shipped null
depends on TASK-IMP-106 · blocks none

TASK-IMP-121: Uninstall must leave the repo as it found it

Summary

A sibling task, TASK-IMP-126, has shipped (status done) and made uninstall the inverse of install for three artifacts: the dangling skill symlinks across every agent family, the two dead MCP registrations, and the pre-commit hook's leading blank separator. This task is what 126 did NOT fix - the uninstall-completeness remainder - and it leads with a live DATA-LOSS bug 126 never touched.

uninstall.sh:141-142 rmdirs .agents/skills and then .agents for emptiness alone, with no ownership test, so an operator's PRE-EXISTING empty .agents/skills is destroyed on uninstall - reproduced end to end at HEAD, and .agents with it when CYBEROS_AGENTS filters the antigravity pointer off. Two smaller defects are byte-level and remain after 126: uninstall's .gitignore strip (:106, still the old sed) leaves the one separator byte install writes, and 126's shipped hook-strip - byte-exact for a newline-terminated hook - still adds one byte to a foreign hook that has no trailing newline. So a directory an operator owns does not come back at all, and two operator files do not come back byte-identical.

Problem

Measured at HEAD (payload built with build.sh; real install.sh/uninstall.sh; GNU awk 5.1.0, bash 5.1). TASK-IMP-126 already removed every dangling skill symlink (its §1.2, uninstall.sh:122-166), both MCP registrations (its §1.1, uninstall.sh:168-186), and the hook separator for a newline-terminated hook (its §1.3, the awk at uninstall.sh:87-93). Those three are OUT of this task. What is left is one data-loss bug and two byte-leaks.

1. A directory an operator owns is destroyed - the lead, and a live data-loss bug. uninstall.sh:141-142 runs rmdir "$root/.agents/skills" then rmdir "$root/.agents", each 2>/dev/null || true. Neither carries an ownership test; rmdir's refusal on a non-empty directory is the ONLY guard, and it does not fire when the directory is an operator's and happens to be empty. Reproduced end to end: a fresh git repo with an operator's pre-existing empty .agents/skills, default install (which drops three symlinks into it), then uninstall - the symlinks are removed and .agents/skills is rmdir'd GONE. In the default config .agents survives, but only incidentally: the :597 antigravity pointer writes .agents/rules/cyberos.md, no uninstall path removes it, so .agents is never empty and :142 no-ops. Filter antigravity off (CYBEROS_AGENTS=agents) and find .agents after uninstall returns nothing - the operator's .agents goes too. This is a casualty a byte-level harness that never seeds a pre-existing .agents/skills cannot see.

Install already records ownership, one directory too low. The discipline uninstall uses everywhere else is: read the artifact, decide from what it says. For a copy-fallback skill dir that discipline is a .cyberos-owned marker: install.sh:684 writes one into each .agents/skills/<cmd> copy dir, and uninstall.sh:130/:161 read it to remove only installer copies. But that marker sits on the CHILD skill dir ($_sdest, install.sh:662), never on the parent .agents/skills or on .agents - and in a default (symlink) install it is not written at all. So the two directories :141-142 can delete carry no proof of who created them, which is exactly why the rmdir is blind. The fix has two halves: install writes a parent marker when - and only when - install itself creates the directory; uninstall then removes those two dirs only when they carry that marker and are empty.

The create-test must precede the mkdir, and each level needs its own. Measured: mkdir -p on an already-existing directory exits 0 silently, so a marker written after mkdir -p "$d" would stamp an operator's pre-existing directory as ours - the current bug inverted, and worse, because it would manufacture a licence to delete. And mkdir -p a/b creates both levels, so a marker in .agents/skills proves nothing about .agents; compounding it, .agents has a SECOND creator at install.sh:575 (the :597 antigravity pointer, mkdir -p on dirname .agents/rules) which runs before :669's mkdir -p "$root/.agents/skills". Each directory must be tested for its own prior existence before the first line that could create it.

Two dirs get markers; five get their errant rmdir removed. grep -n rmdir uninstall.sh returns :141, :142, :165 and :226 (:226 is the BRAIN-restore tmpdir, unrelated). :141-142 are the .agents pair, which install CREATES and wants to reclaim cleanly, so they get the marker mechanism (§1.1-1.3): delete only what we proved we created. :165 is the per-family loop's rmdir "$root/$_sd" over the five native channel parents (.claude/skills, .grok/skills, .commandcode/skills, .codex/skills, .opencode/skill) - added by TASK-IMP-126 and absent pre-126, and reproduced destroying an operator's pre-existing empty .claude/skills end to end at HEAD. Those parents were never reclaimed pre-126, so the fix is not a marker but the REMOVAL of the blind rmdir (§1.6), restoring the pre-126 leave-in-place. A marker there would be a file nothing reads - a new permanent leftover; removing the rmdir is both smaller and the pre-126 status quo. So the same data-loss class (rmdir on emptiness with no ownership test) is closed at all four sites: the .agents pair by marker gate, the five native parents by deleting the rmdir IMP-126 added.

2. The .gitignore separator leak. install.sh:756 writes [ -s "$gi.cyberos.tmp" ] && printf '\n' before the managed block, and uninstall.sh:106's sed range starts AT the # >>> cyberos marker, so the separator above it survives. IMP-126 fixed the hook this way but not .gitignore. Measured by replaying real install+uninstall on a pre-existing newline-terminated .gitignore: 20 -> 21 bytes, holding at 21 across three cycles (install.sh:742-754's awk trims the trailing blank run on re-install, so the leak is one-time, not accumulating). Still a mutation of an operator's file, and the fix belongs to the strip.

3. The hook, only for content with no trailing newline. IMP-126's awk (uninstall.sh:87-93) is byte-exact for a foreign hook that ends in a newline (measured 34 -> 34), which is what its arm t_hook_strip_byte_identical_across_cycles proves. It is NOT byte-exact for a hook with no trailing newline: measured 33 -> 34 bytes (+1). Install's cat >> appends its leading separator directly after the last unterminated line, so there is no blank LINE for the awk to hold, and awk's print re-terminates that last surviving line. The information is not destroyed - install's append stays invertible - but only a byte-level strip (delete the block plus the one newline byte immediately before the # >>> marker) recovers it; the line-oriented awk cannot.

Proposed Solution

Four independent, testable changes, each grounded in a re-verified HEAD citation.

Gate the container rmdir on a parent marker install writes on creation (the data-loss fix). install.sh writes a .cyberos-owned marker into .agents/skills, and separately into .agents, when and only when install itself creates that directory - deciding creation by testing the directory's own existence BEFORE the mkdir -p that could create it (:669 for .agents/skills; :575-via-:597, which runs first, for .agents). uninstall.sh:141-142's unconditional rmdir is replaced by: remove the parent marker, then rmdir only if the directory now carries our marker and is empty; a directory with no marker is kept and reported as kept. Where install created the directory, uninstall takes it back and leaves nothing behind; where the operator's directory predated install, it survives. This is TASK-IMP-094's mechanism applied one directory up.

directoryproof it is oursconsequence
operator's pre-existing .agents/skillsno parent markerkept + reported
install-created .agents/skillsparent marker (§1.1)marker removed, then rmdir if empty
operator's pre-existing .agentsno parent markerkept + reported
install-created .agents (marker kept dir)parent marker (§1.1)marker removed even though the dir is kept (rules pointer survives)

Delete the native-parent rmdir IMP-126 added (the other half of the data-loss close). uninstall.sh:165's per-family rmdir "$root/$_sd" removes an operator's pre-existing empty .claude/skills (and the other four native parents) once our managed link is cleared - the same blind-emptiness deletion as :141-142, introduced by IMP-126 and absent pre-126. Remove that one line (§1.6), so an emptied native channel dir is left in place like any other inert container. No marker is written or read there: pre-126 uninstall never reclaimed these, so leaving them is both correct and the status quo, and it keeps IMP-126's link removal intact.

Strip the .gitignore separator with the block, by the byte. The .gitignore strip must also delete the single \n byte install writes immediately before the # >>> cyberos marker (install.sh:756), restoring a pre-existing newline-terminated .gitignore byte-identical across any number of cycles. Same look-behind shape IMP-126 already uses for the hook, ported to the .gitignore strip.

Make the hook strip byte-exact for a hook with no trailing newline. Refine the strip so a foreign hook is byte-identical to its pre-install content even when that content has no trailing newline: delete the block together with the one separator byte and do not re-terminate the last surviving line. The newline-terminated case (already byte-exact, IMP-126) must stay byte-exact.

Alternatives Considered

Success Metrics

Scope

In scope: uninstall.sh:141-142's container rmdir, now gated on a parent .cyberos-owned marker with the marker removed before the rmdir; install.sh's create-detection and the two parent markers it writes (.agents/skills at :669; .agents at :575-via-:597); uninstall.sh:165's per-family rmdir over the five native channel parents, REMOVED so an operator's pre-existing empty native dir survives; uninstall.sh:106's .gitignore strip, made byte-exact for the separator; uninstall.sh:87-93's hook strip, made byte-exact for a no-trailing-newline hook; and new arms in test_install_hygiene.sh for each.

Out of scope / Non-Goals

Dependencies

depends_on TASK-IMP-106 (status done): 106 landed the uninstall summary, the _removed_list/_kept reporters, and the t22 arm on the same uninstall.sh + test_install_hygiene.sh surface this task edits, so 106's file state is the base these edits build on. The pre-126 justification for the dependency - "121 must re-point 106's t22" - is RETIRED: re-verified at HEAD, t22 is self-relative and unaffected by this task, so no re-point is authored. Shares surface with TASK-IMP-094 (the .cyberos-owned marker mechanism this task lifts one directory up) and TASK-IMP-126 (the sibling that shipped the symlink/MCP/hook-core; this task is its remainder).

AI Authorship Disclosure

1. Description (normative)

2. Acceptance criteria

3. Edge cases

Audit

TASK-IMP-121 - audit record

§0 - Provenance, and why this file had to be reconstructed

This page was rebuilt on 2026-07-18. Read this section before trusting any finding id on it.

The 6/10 audit was written into a COMMIT MESSAGE and never into this file, which sat at the 4/10 version (6a146869) while two rewrites happened against findings no one could open. The defect is recorded as NEW5-007 on TASK-IMP-122's audit page. This file is the remedy for 121's half of it.

§0.1 - What kind of section you are reading

kindsectionswhat you are reading
written at the time§1 (audit 1, 4/10)the auditor's own audit file, committed at 6a146869, preserved verbatim
reconstructed from commit§2 (audit 2, 6/10)the orchestrator's summary of an audit. The audit's own text does not exist.
not an audit at all§3 (the third author's fix)the author's closure claims. Nobody audited this.

§0.2 - The rounds, mapped

sectionwhat it issourcepersisted at the time?verdict
§1audit 1, of the first draft6a146869 (this file)YESFAIL 4/10
§2audit 2, of the first rewrite15894b1e messagenoFAIL 6/10
§3the third author's fix - the current spec.md1f8143cf messagenoNONE. UNAUDITED.

All six ids audit 2 raised - NEW-001 through NEW-006 - resolve to §2. Neither spec.md cites a finding id by name (verified: grep -oE 'NEW-[0-9]{3}|ISS-[0-9]{3}' on both specs returns nothing for 121), so no citation in the task depends on this page. source_decisions refers to the rounds by score - "2026-07-18 audit FAIL 4/10" (§1) and "2026-07-18 audit FAIL 6/10" (§2) - and both now resolve.

§0.3 - A citation hazard a reader will otherwise walk into

Audit 2's clause and AC numbers do not resolve against HEAD's spec.md.

Audit 2 examined the first rewrite: 9 clauses, 9 ACs, effort 5 -> 8 (per 7309cd80). HEAD is the third author's whole-document rewrite: 10 clauses, 10 ACs, effort 8 -> 9. The renumbering between them was never mapped, and §2's citations are preserved as the auditor wrote them rather than silently re-pointed - re-pointing them would be a reconstruction of a mapping nobody recorded.

Concretely: NEW-002 is about "§1.5's six channel dirs" and "AC 5". In HEAD, the six-channel container rule is §1.6 and its test is AC 6; HEAD's §1.5 is the operator-edited-.mcp.json rule and its AC 5 is that rule's test. Reading NEW-002 against HEAD's §1.5/AC 5 will mislead you. The same hazard applies to NEW-001 (§1.6/AC 6/AC 7 as numbered in the first rewrite; the byte-exact strip rule is §1.7 in HEAD) and NEW-004 (§1.3 vs §1.8).

§1 - AUDIT 1: audit of the first draft - FAIL 4/10

Provenance: WRITTEN AT THE TIME. This is the audit file as committed at 6a146869 (2026-07-18). It is the only section of this page that is an auditor's own artefact. Preserved verbatim; nothing below this line in §1 has been edited.

Status: superseded as a verdict, and its findings are resolved - audit 2 (§2) records ISS-001/003/004/005 as DISSOLVED by the frame replacement rather than patched, and all six of ISS-006's citations as fixed. ISS-002's shape RECURRED on a new input as NEW-003 (§2.4).

Audit 1's own header block, as written:

task_id: TASK-IMP-121
audited: 2026-07-18
verdict: FAIL
score: 4/10
issues_open: 6
issues_resolved: 0
template: task@1
audit_rubric_version: audit_rubric@2.0
machine_floor: task-lint clean (exit 0) - run FIRST per TASK-IMP-084. The floor passing is why this
  audit matters: every finding below is a JUDGMENT defect the mechanical check cannot see.
auditor: independent subagent (had not seen the author's reasoning) + author verification of the
  two load-bearing claims against source

§1.1 - Verdict summary

FAIL at 4/10. The four underlying defects are REAL and reproduced by the harness. The spec diagnoses two of them from a false premise, carries the TASK-IMP-118 defect class in its own AC 1, and cites six line numbers that do not resolve. Not a patch: a rewrite against the true mechanism.

§1.2 - Findings (ALL OPEN)

ISS-001 (CRITICAL) - the premise ".cyberos is removed" is false in the default path

§1.1/§1.3 are written against "the removed .cyberos/". VERIFIED against source: uninstall.sh:151 runs rm -rf "$CY" and :155-158 then runs mkdir -p "$root/.cyberos/memory" + mv "$KEEP_BRAIN_STASH" "$root/.cyberos/memory/store". .cyberos/ survives with BRAIN inside unless CYBEROS_UNINSTALL_KEEP_BRAIN=0. The harness's own fresh-git run listed ./.cyberos/memory/store/* among the leftovers - the author had the evidence and wrote the clause against the opposite state.

ISS-002 (CRITICAL) - AC 1 asserts a different predicate than §1.1 (the TASK-IMP-118 defect)

§1.1 verb: a symlink "may remain [only if its] readlink target [does not resolve] inside the removed .cyberos/". AC 1 asserts: "zero unresolvable targets". These diverge BOTH ways:

  • false negative: with BRAIN restored (ISS-001), a symlink into a SURVIVING .cyberos/ path resolves fine -> passes AC 1 -> violates §1.1.
  • false positive: an operator's own broken symlink pointing OUTSIDE .cyberos/ is unresolvable -> fails AC 1 -> but §1.2 requires it be KEPT. AC 1 and AC 2 contradict on that input. The author wrote both the clause and its test, and the test is weaker. This is precisely §15.2.

ISS-003 (CRITICAL) - §1.3/§1.4 are undecidable; the enabling mechanism is rejected in-spec

Both condition on "if install created it". Nothing on disk records that: install.sh tracks creation in a shell variable that dies with the process; the only ownership marker (.cyberos-owned) is written ONLY into skill copy dirs, never onto .gitignore or .mcp.json. The spec's Proposed Solution permits only "readlink target inside .cyberos/, or our marker" - readlink is meaningless on a regular file and no marker exists - while Alternatives REJECTS "Track install's creations in a manifest" as scope creep. The spec forbids the only mechanism that would make its own clauses decidable.

ISS-004 (CRITICAL) - §1.4 contradicts §1.7 and the zero-casualties guardrail

An operator who ran touch .gitignore before install is byte-indistinguishable from install's : > "$gi". After block-strip the file is empty; §1.4 says it MUST NOT survive -> uninstall deletes an operator's file -> §1.7 ("No path present before install may be missing") violated. §1.7 admits no exception and §3 does not cover the case. Same hazard for .mcp.json: install points operators at .cyberos/mcp/README.md for hand-registration, so an operator-authored .mcp.json is a supported path.

ISS-005 (MAJOR) - section 6 already ACCEPTS the dangling links; the spec frames it as oversight

VERIFIED: uninstall.sh:162 reads # 6. skill symlinks into .cyberos (dangling) - leave dirs; operator cleans. The dangling is KNOWN and DELIBERATE in source. The spec's Problem section says four channels "are never looked at" - true of the code path, false of the intent. This is a recorded decision to OVERTURN with reasons, not a gap to fill. The operator approved overturning it (2026-07-18) on the readlink-proves-ownership argument, which stands - but the spec must argue against section 6 explicitly rather than not notice it.

ISS-006 (MAJOR) - six citations do not resolve

  • :70 cited as "(block strip)" and as the proven newline leak -> :70 is echo " stripped cyberos block from pre-commit". The sed is :68. This is the spec's single most load-bearing citation and it points at the echo, not the mechanism.
  • :81 cited as the .gitignore strip -> :81 is the echo; the strip is :80.
  • :114 cited BOTH as the ship-tasks exemption AND as the readlink test it exempts from. One line cannot be both. The [ "$_sc" != "ship-tasks" ] guard is :112.
  • :111 cited as an exemption whose rationale is "avoid clobbering operator files" -> :111 is a comment; it records no such rationale.
  • :105-107 cited as a prune precedent -> it is the "kept unmarked skill dir" echo branch. The actual prune is :118-119 and has NO install-created check.
  • :92-93 resolves but is a comment naming 3 paths, cited to cover 7.

§1.3 - Verified accurate (credit)

  • uninstall.sh never mentions mcp/codex/grok/commandcode/opencode - grep returns zero. The five-channel and dead-registration findings are REAL.
  • The +1-newline-per-cycle diagnosis is mechanically correct and fixable from uninstall.sh alone: install appends \n# >>> cyberos-status-hook v2, and :68's sed range starts AT the marker, so the blank separator above it survives.
  • 121's reading of TASK-IMP-106 AC 3 is accurate; depends_on: [TASK-IMP-106] correctly serialises.
  • TRACE-003 passes: test_install_hygiene.sh exists and is declared in modified_files.

§1.4 - Required before re-audit

Rewrite §1.1/§1.3 against the surviving-.cyberos reality; re-derive AC 1 to assert §1.1's actual predicate; resolve ISS-003 by either scoping the creation-manifest IN or narrowing the clauses to what readlink/marker can decide; reconcile §1.4 with §1.7; argue against section 6 explicitly; fix all six citations. Residual noted: after 121 re-points t22, TASK-IMP-106 §1.5 has no test left that verifies it, and a test named t22_uninstall_behavior_unchanged whose baseline moved asserts the opposite of its name.

§2 - AUDIT 2: audit of the first rewrite - FAIL 6/10

Provenance: RECONSTRUCTED FROM COMMIT 15894b1e (2026-07-18); no contemporaneous audit file exists. What follows is the orchestrator's summary of an independent audit. The auditor's own text is gone. NEW-001 carries a note that the orchestrator verified it against source independently ("I verified this myself, it is exact"); the other five carry no such note.

The subject was the FIRST rewrite (9 clauses / 9 ACs, 7309cd80), written by a fresh author who had written none of the failing draft. It is not HEAD. See §0.3 before resolving any clause number below.

§2.1 - Verdict summary

FAIL 6/10 (from 4/10). The frame-replacement does real work: ISS-001/003/004/005 are DISSOLVED not patched, all six inherited citations fixed, every inherited number reproduced exactly. It fails on what the author ORIGINATED.

§2.2 - NEW-001 (CRITICAL) - the impossibility proof is FALSE

The orchestrator's note records this as independently verified: "I verified this myself, it is exact."

§3 claims an operator hook with no trailing newline "cannot be restored byte-exact - the information is destroyed at append time and no uninstall-side rule inverts it". FALSE. The information survives as the presence/absence of the blank line above the marker, and install ALWAYS prepends exactly one \n, so the byte before the marker is always ours. Measured:

line-oriented strip (the spec's):  6B -> 7B  DIFFERS
byte-oriented rule (awk, no perl): 6B -> 6B  BYTE-EXACT   <- inverts it
same rule, newline-terminated:     7B -> 7B  BYTE-EXACT

A LEVEL CONFUSION: the author reasons in LINES, and in line-space the no-newline case has no blank line to consume - so they generalised "my line-oriented rule cannot" to "no rule can".

§1.6 is NARROWED on that false proof, which excuses a live §1.8 violation (an operator's hook permanently mutated +1 byte) and scopes AC 6/7 so the suite can never catch it.

§2.3 - NEW-002 (MAJOR) - AC 5 is UNSATISFIABLE

§1.5 requires each of §1.1's six channel dirs to survive and forbids removing a container for emptiness. uninstall.sh:118 rmdirs .agents/skills - the FIRST of the six - for exactly that. Verified. §3 calls it "mild tension" and re-scopes §1.5 to five IN PROSE ONLY; the clause and the AC still say six.

Aggravating: an operator's pre-existing EMPTY .agents/skills is deleted - a §1.8 violation AC 8's four cases miss.

§2.4 - NEW-003 (MAJOR) - AC 1 and AC 2 contradict: ISS-002's exact shape on a new input

§1.1's chained-form pattern widens the hazard from the 3 names install writes to EVERY entry in the dir, destroying the ours-by-construction justification while keeping the pattern. §3's "carried forward unchanged" is false.

ISS-002 (§1.2) is the same defect class - clause and test diverge, author wrote both - recurring after being dissolved once.

§2.5 - NEW-004 (MAJOR) - §1.3 vs §1.8: ISS-004's shape relocated

From container-DELETION to content-MUTATION. The §1.3/§1.4 split fixed the drafting, not the substance: an operator's hand-registered .mcp.json byte-identical to :685's form triggers §1.3, which then mutates a path present before install.

§2.6 - NEW-005 (MINOR) - a SEVENTH mis-citation, ORIGINATED

:684 is an if-guard; the summary is :689 - under a disclosure reading "every line number was re-verified". The six inherited citations were all genuinely fixed; the seventh is the author's own.

§2.7 - NEW-006 (MINOR) - the receipt rejection's conclusion stands but one universal is false

Under CYBEROS_COPY_SKILLS=1 the copy is decidable from a receipt and from nothing else - which is §3's own conceded gap. The conclusion (reject the receipt) survives; the universal supporting it does not.

§2.8 - THE PATTERN, ACROSS SEVEN AUDITS AND THREE AUTHORS

Recorded by the orchestrator alongside this audit, because 121 is the round that made it legible:

round 3 (me): patched what the audit NAMED, did not re-read     -> 4 findings survived verbatim
round 4:      re-read everything, false NUMBER of their own      -> 1525
121 rewrite:  fixed every citation, false REASONING of their own -> the impossibility proof

One rule underneath all three: AUTHORS DO NOT CHECK WHAT THEY ORIGINATE. An inherited claim carries a provenance that invites scrutiny; an originated one arrives already believed. It is TASK-IMP-118's defect class one level up - there, an author's test asserts something weaker than their own clause; here, an author's evidence never gets tested at all. Both are the same failure: the author is the wrong reader of their own work.

This became TASK-IMP-124. Note that 124 rejected rotate-the-author using the evidence offered FOR it: both fresh-author rounds shipped an originated falsehood. "Rotation changes who INHERITS; it does not make anyone check what they ORIGINATE."

§3 - THE THIRD AUTHOR'S FIX: CLAIMED CLOSED, UNAUDITED

Provenance: RECONSTRUCTED FROM COMMIT 1f8143cf (2026-07-18); no contemporaneous file exists.

THIS IS NOT AN AUDIT AND THIS SECTION RECORDS NO VERDICT. It is the third author's account of closing audit 2's six findings, relayed through the orchestrator's commit message. HEAD's spec.md is this rewrite. No independent auditor has read it.

§3.1 - The status, stated exactly

The six findings NEW-001..006 are CLAIMED CLOSED. The fix is UNAUDITED.

  • No third audit exists. Not in this file, not in any commit message, not anywhere.
  • The closure claims are the author's own - which is precisely the category audit 2's own §2.8 identifies as the least-checked thing in any document. The rule the round established applies to the round's own output.
  • score_history therefore stops at 6/10. The 6/10 is audit 2's verdict on the draft that PRECEDED this fix. It is not HEAD's score, and this file does not assign HEAD one. A reader who needs a verdict on HEAD must commission an audit; there is nothing here to read one off.

Two things partially corroborate the claims, and neither is an audit:

  1. The orchestrator reports reproducing NEW-001's disproof independently: "line-strip 6B->7B DIFFERS, byte-strip 6B->6B EXACT, controls 7B/9B exact." That is one finding's evidence re-measured, not a review of the document.
  2. The machine floor passes: task-lint clean, 10 clauses / 10 ACs / 9h. Audit 1's own header says why that is worth little here: "The floor passing is why this audit matters: every finding below is a JUDGMENT defect the mechanical check cannot see."

§3.2 - What the third author claims, per 1f8143cf

All six findings closed. Specifically recorded:

  • NEW-001. Reproduced the disproof independently and retracted §3's impossibility. Then found the impossibility IS REAL - on the OTHER FILE: install.sh:733-746's awk collapses *.log\n (6B), *.log\n\n\n (8B) and *.log (5B) all to 6B. Three distinct pre-install states, one post state. A true pigeonhole, install-side, correctly scoped out.
  • It tested its own new rule and found its limit - operator content BELOW our block with no trailing newline: want 17B, got 18B - and recorded it rather than papering it. (This is the §2.8 rule being applied by an author to their own work, which is the first instance of it in the session.)
  • §1.9's carve-out was FORCED, not chosen: §1.9 contradicted §1.1, not only §1.4 - an operator's pre-existing symlink whose target names our machine is a path present before install that §1.1 removes. The 6/10 audit found the §1.4 collision and MISSED the §1.1 one.
  • effort 8 -> 9.

§3.3 - OPEN: needs an operator verdict, not an auditor's

Bringing uninstall.sh:118-119 in scope - keeping six channel dirs rather than scoping §1.6 to five - is the author's decision and is NOT covered by the recorded 2026-07-18 PLAN gate. Flagged by the author in the spec's own AI Authorship Disclosure. This is a scope question for the operator and an audit cannot settle it.

§3.4 - Also filed against this round

1f8143cf records a finding about TASK-IMP-118, raised while filing TASK-IMP-124 and belonging to neither 121 nor 122: TASK-IMP-118 declares modified_files: tools/install/docs-tools/templates/ task-audit/RUBRIC.md - WHICH DOES NOT EXIST (verified). TRACE-003 covers test paths only, so an originated citation about the author's own cone is checked by nobody. Recorded here only so it is not lost with the commit message; it needs its own task and is not 121's.

§3.5 - Required before this task can be trusted

  1. Audit HEAD. The six closures are unverified. This is the whole of what is missing.
  2. Get the operator's verdict on §3.3 - :118-119's scope is outside the PLAN gate.
  3. Note for whoever audits: §2's clause numbers do not resolve against HEAD (§0.3). The mapping from the 9-clause draft to HEAD's 10 was never recorded and is not reconstructed here.

Not promoted; no BACKLOG row.