Preserve PR561 lifecycle fixes and retire source-host evidence

[seathatflowsinourveins]

Scope

Preserve PR561's hook ownership boundaries, explicit no-project Serena parity and opt-in saved-workflow installer. Retain its source-host facts in a dated retirement record and carry 17 anti-pattern rows after comparing all 19 candidates against current main; two duplicates stay represented by main's existing rows.

• Base commit: 7d0174168d478d8e3a01602db2138905b6130f8a (main). Head: 3758ce5218321a421476248e97b70034a3c0ab2f.
• Lane: lane:foundation.
• Owned paths: the settings merger, native-token fixture and profile installer with their three test modules; two bounded bootstrap edits; anti-pattern table rows; the new retirement record; four unchanged Serena preimage files; evidence file registrations in the last commit.
• Commits: 86b98f24e ("Port PR561 hook isolation, Serena parity and opt-in workflows", 13 paths) and 8fea93403 ("Refresh evidence bindings for the PR561 lifecycle port", manifests/evidence.json only), then the review repair round: c1ae00392 ("Mark PR561-only references in the ported anti-pattern rows": docs/harness-defaults.md, the retirement record and tests/test_install_claude_profile.py) and 3758ce521 ("Refresh evidence bindings after the PR561 review repair", manifests/evidence.json only, the last commit).
• Registry: 8fea93403 starts from main's registry at the base and registers the five new files and the eight changed files main already lists; 3758ce521 re-registers the three files the repair changed. The component-matrix and grand-list --check runs passed in both rounds, so no report was regenerated. The other 9,562 rows stay byte-equal to main's, in main's order. No receipts[] or convergence_records[] entry is added.
• Main is at e0c329ae98efb6ac5e11a49220b69ec6eddbb658 (macOS CI scope: validate-macos stays required, running the full suite for Mac-relevant changes, only changed test modules for test-only changes, and skipping otherwise (decision record, receipt, tests) #677), past the base. This PR is not rebased onto it. The merge-tree landing check of head 3758ce52 against it returns LANDABLE: main's 34 changed paths share only manifests/evidence.json with this PR, and that file's foreign rows stay equal and in order. The merge-time hot-file rebase covers it.
• Keep this PR a draft until the Codex root lane's exact-head read and all eight required contexts pass at the final head. The custody contract requires no other-lane ack.

SOTA sources

• akitaonrails/ai-memory v2.5.2, commit 7580b74d0fb9d14a6d949dc92f5ea8bb7feb3c83: install_hooks.rs L1583 ownership classifier, L1619 event overlay, and uninstall.rs L825 command signatures. All three originals were reread; tag object af8c6820bbbcca0b8d3604d7b2d59c0e0263ea85 resolves to that commit.
• ai-memory 2.4.2 reproduction pin a0ca8d1a5fbd5920799411fa891fe6d49c90efc1: install_hooks.rs L1513–1565.
• oraios/serena c6fbd1c5932df2494ffa0020af5a9fbe80b82143: cli.py L369–383, agent.py L699–701, L839–848.
• Claude's save-workflow layout and target-symlink rule, with main's reviewed workflow SHA256SUMS at the base. Workflow bytes remain unchanged; installation is opt-in.
• PR561 head fe4729d98b16fb71445373b6da4ea07da8b4b8cf is the origin of every ported hunk and retained evidence byte. Main's narrow installer options and default guard/agents/MCP steps are preserved.
• Evidence registration/generation follows this repository's hot-file protocol, register_file, and maintained component-matrix/grand-list generators. No receipt or convergence-record entry is added.

Evidence-class table

Claim	Evidence class	Command / receipt
Ownership classifier, ancestor activation and saved-file layout justify the integrations	source_review	Exact pinned originals and official workflow documentation above
Ported code and the unchanged new-WSL integration tests pass at the head	local_integration	Four-module command: 326 tests, 4 skips, exit 0; all 159 new-WSL tests pass unchanged
Canonical template hook entries stay separate, and existing mixed entries split into contiguous runs	local_integration	tests.test_apply_claude_settings: 25 tests, exit 0. Discriminating control below: the three hook-ownership tests fail against the base merger
Explicit workflow selection, default no-access behavior, conflicts, readback, rollback and scoped removal pass	local_integration	WorkflowInstallTests: 15 tests, no skips, exit 0. Discriminating control below: the default no-access test fails once the default steps include workflows
Four Serena files retain the historical no-project observations	native_proven (source host, 2026-09-30; not rerun)	The four committed blobs are identical to fe4729d's; both stdout SHA-256 values equal the observation's file_sha256
Historical f03fb80 CI returned 259 commands / 82 checks	source_review of retained native execution	Run 36792680424, job 110148966262 (native-token-tools, success, head f03fb80) metadata and returned log reread; no job rerun
Evidence, catalog, host-receipt, anti-pattern-table and generated-report structure passes at the head	structural_validation	validate.py (9,575 hashed files), all-recorded convergence (29 valid records), host receipts (182), catalog validator, component matrix (32 rows), grand list (32 layers, 66 winners), ecosystem check, evidence-manifest order check (9,575 rows), workflow checksums (14 OK) and the docs-consistency module (39 tests, 1 skip), each exit 0
The head merges cleanly and the registry keeps every foreign row	structural_validation	Merge-tree landing check: clean three-way merge, foreign rows equal and in order, 13 PR-owned rows, sorted files[] (9,601 merged rows); LANDABLE against e0c329ae at head 3758ce52

Discriminating controls

Each control ran on a git archive copy of head 3758ce52 with Python 3.13.15, nice -n 19 and a TMPDIR under /var/tmp. <repo> stands for the copy's root. Only the named file was changed in the copy, and the repository was not edited.

Claim	Passing run (as merged)	Condition removed	Failing run
A default run never reads or creates workflows	tests.test_install_claude_profile.WorkflowInstallTests.test_default_profile_never_reads_or_creates_workflows: exit 0, Ran 1 test, OK	tools/adoption/install_claude_profile.py line 458 set to steps = args.only or ["guard", "agents", "workflows", "mcp"]	exit 1, FAILED (failures=1): AssertionError: default run accessed workflows: <repo>/examples/claude-native/workflows/readiness-audit.js
Template entries stay separate; mixed entries split	MergeSettingsTests test_absent_and_empty_matchers_keep_their_own_groups, test_existing_mixed_entry_is_split_without_changing_hook_values_or_order and test_same_matcher_keeps_native_memory_and_carrier_entries_separate: exit 0, Ran 3 tests, OK	tools/adoption/apply_claude_settings.py replaced by the base 7d0174168 merger (blob f07168fc)	exit 1, FAILED (failures=3): AssertionError: 2 != 4, AssertionError: 1 != 3, and a Lists differ on the canonical entries

The same two controls also failed and passed identically on a copy of the reviewed head 8fea9340.

Fresh CI starts with the push of 3758ce52; no result for it is claimed here. At the reviewed head 8fea9340, every completed check passed, including native-token-tools; validate-macos was cancelled at 2026-10-04T01:24:03Z because this push superseded that run (adoption-bootstrap.yml sets cancel-in-progress for its pull-request concurrency group). Native-preimage replay and historical CI reads establish no new provider run. Exact source-host activation clock times and the eight-Claude-hook count retain custody-contract provenance rather than a stronger independently observed claim.

Local commands run

Every command ran at head 3758ce52, outside any sandbox, with nice -n 19 and a writable TMPDIR under /var/tmp. origin/main was e0c329ae; the merge-base stays the base 7d0174168. An earlier sandboxed preparation run is not acceptance evidence: its read-only /var/tmp and cgroupfs caused failures, and none of its counts are carried here.

$ python3 -m unittest tests.test_apply_claude_settings tests.test_install_claude_profile tests.test_native_token_ci tests.test_new_wsl_client_config tests.test_windows_terminal_defaults tests.test_currency_due_notice tests.test_adoption_bootstrap tests.test_adoption_bootstrap_macos -v
exit 1: Ran 631 tests in 101.691s; FAILED (failures=1, skipped=36)
The one failure is pre-existing on main (classification below).

$ python3 -m unittest tests.test_apply_claude_settings tests.test_install_claude_profile tests.test_native_token_ci tests.test_new_wsl_client_config -v
exit 0: Ran 326 tests in 41.357s; OK (skipped=4)

$ python3 -m unittest tests.test_new_wsl_client_config
exit 0: Ran 159 tests in 38.167s; OK

$ python3 -m unittest tests.test_install_claude_profile.WorkflowInstallTests -v
exit 0: Ran 15 tests in 0.642s; OK

$ python3 -m unittest tests.test_apply_claude_settings -v
exit 0: Ran 25 tests in 0.027s; OK

$ python3 -m unittest tests.test_adoption_docs_consistency
exit 0: Ran 39 tests in 11.612s; OK (skipped=1)

$ python3 -m unittest
Not run locally. Per the custody contract's publication amendment, CI's validate job runs the full suite on this head.

$ python3 scripts/validate.py
exit 0: 69 components, 9575 hashed files, 4 profiles, 193 receipts; passed

$ python3 scripts/host_receipts.py validate
exit 0: 182 receipts; passed

$ python3 scripts/validate_catalogs.py
exit 0: catalog structure and evidence classes validated; source claims and native executions were not rerun

$ python3 scripts/validate_convergence.py --all-recorded --root . --json
exit 0: valid true, 29 records

$ python3 scripts/component_matrix.py --check
exit 0: 32 rows, checked

$ python3 scripts/new_host_grand_list.py --check
exit 0: 32 layers, 66 winners, passed

$ python3 scripts/build_ecosystem.py --check
exit 0: passed; the existing architecture pin drift stays reported

$ python3 scripts/evidence_manifest.py --check
exit 0: 9575 files, passed

$ (cd examples/claude-native/workflows && sha256sum --check --strict SHA256SUMS)
exit 0: all 14 files OK

$ git diff --check origin/main...HEAD -- . ':!evidence/artifacts'
exit 0, no output

$ git diff --check origin/main...HEAD
exit 0, no output

$ convergence rebinding query (contract step 19) over main's convergence_records at e0c329ae
exit 1, empty output: no record cites a changed path, so nothing is rebound

$ four Serena files against the fe4729d9 blobs and the observation's file_sha256 values
exit 0: all four identical; both stdout hashes match

$ merge-tree landing check, origin/main e0c329ae against head 3758ce52
exit 0: LANDABLE (clean three-way merge; 14 merged-vs-main paths, none outside the PR; no overlap with main's 34 changed paths besides the registry; foreign rows equal and in order; 9601 merged rows sorted)

$ git push (pre-push hook: three registry tests on 3758ce52)
exit 0: Ran 3 tests; OK. The gitleaks pre-commit hook passed on both repair commits.

Failure classification

Test	Class	Main baseline
tests.test_windows_terminal_defaults.OverlayTests.test_the_installed_client_knows_no_notification_type_without_a_decision	pre-existing: fails identically on main. It reads this host's installed Claude Code client, which reports auth_storage_failure with no DECISIONS entry.	Same eight-module command at 7d0174168: exit 1, Ran 613 tests, FAILED (failures=1, skipped=36), the same assertion. The single test on a copy of current main e0c329ae: exit 1, the same Lists differ: ['auth_storage_failure'] != []

The 36 skips have identical reasons on main and at the head: 32 macOS bootstrap prerequisites (17 without a bash 3.2 binary, 14 without PyYAML, 1 without shellcheck), 3 PyYAML frontmatter cases and 1 unset Context Mode security module path. This PR adds no skip and introduces no failure.

Decision record

docs/decisions/2026-10-03-pr561-token-lifecycle-retirement.md records alternatives, upstream pins, all 139 source paths (13 ported / 126 retired), unique source-host facts, explicit non-claims, two dropped duplicate rows, the ten current-enforcer and seven historical-only ported rows, and the comparison/acceptance conditions that reopen each retired category.

Host evidence

No evidence/hosts files or platform-status claims change. The four copied artifacts retain their original source-host/date scope.

Checklist

• No GitHub Actions workflow is added or changed.
• Workflow script bytes and their checksum manifest remain unchanged.
• Main's pins, templates, narrow installer APIs and default steps remain.
• File registrations use the maintained protocol; receipts[] and convergence_records[] remain unchanged.
• Peer-owned paths and the stale PR's branch/head are preserved.
• No credentials, raw conversations or client configuration quotations are published.
• No paid hosting or new billing surface is introduced.
• Based on main 7d0174168, not rebased onto e0c329ae (the landing check covers it); reran the contract's local checks at head 3758ce52 with a writable TMPDIR; the full suite is left to CI's validate job.
• Fresh native-token-tools and all eight required contexts pass at the final head.
• Independent Claude Opus/max exact-head review (of 8fea9340, verdict repair) and its one repair round are recorded; the dispositions are in the review-record comment.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Claude session native-agent-stack-0c: wsl-architecture-design: merge_settings, imported by new_wsl_client_config.py, now keeps canonical template entries separate and splits existing mixed entries into contiguous runs without changing hook values or order. test_new_wsl_client_config passed unchanged at 8fea934 (159 tests). This is informational, not an acknowledgement gate.

[seathatflowsinourveins]

Claude session native-agent-stack-0c: review record and repair round for this port.

Review. An independent, read-only Claude Opus 5.5 review at max effort, run as a separate headless session, read head 8fea9340375b. Verdict: repair, with 0 blocking, 2 should-fix and 4 minor findings.

What the reviewer checked (as it reported them, one line each):

• Contract content against fe4729d: apply_claude_settings.py differs only by the contracted v2.5.2 docstring. Two test modules are identical to the source. native_token_ci.py keeps the ai-memory 2.4.1 pin and changes only the comment block and the --project-from-cwd argv; the SERENA_* hashes are unchanged. The installer keeps default steps guard/agents/mcp and the narrow options, bootstrap.md has exactly 2 hunks, and the bootstrap scripts are unchanged.
• Byte identity: the four serena-no-project blobs equal fe4729d's (4f6d3406, 3bf7f69b, f2313641, 7d5cc8c5), and both stdout SHA-256 values equal the observation's file_sha256.
• Paths: 14 files, all inside the allowed paths, and no forbidden or owner path is touched. 86b98f24e holds the 13 content paths and 8fea93403 only manifests/evidence.json. Qualify native Claude/Codex token profile and memory lifecycle #561 is still open, draft, head fe4729d.
• Anti-pattern table: of 19 source rows, 17 were added and 2 dropped as duplicates (both listed in the record). No existing row changed, and the base table is an ordered subsequence of the head. The 6/11 split became finding 1.
• Retirement record: all 139 of 139 Qualify native Claude/Codex token profile and memory lifecycle #561 paths are listed (126 retired, 13 ported) with matching blob URLs. The non-claims are verbatim and the facts match the fe4729d blobs. Run 36792680424 rests on the custody chronology plus a GitHub reread.
• Upstream citations (ai-memory L1583, L1619, uninstall.rs L825 and a0ca8d1 L1513; Serena cli.py L369-383 and agent.py L699-701, L839-848) match the copies the builder fetched. The reviewer did not fetch them independently.
• Evidence policy: historical runs stay labelled historical, and no replay or local test is presented as a new run or as upstream acceptance. The missing controls and class labels became findings 2 and 5.
• Registry: rows went from 9,570 to 9,575, with exactly 5 added and 8 changed for the port's 13 files, all recomputed OK. The other 9,562 rows equal main's, in main's order. receipts[] and convergence_records[] equal the base. evidence_manifest.py --check exited 0.
• Merge claim: main e0c329ae changed 34 paths and shares only manifests/evidence.json with this PR. A three-way merge of that file is clean, the merged files[] is sorted with no duplicates, and both sides' rows are kept.
• Privacy: 1,033 added lines and 129 body lines show no user name, home path, private file name, email address or token pattern. The only hit is the public distro name inside a non-claim the contract requires.
• Reproduced at 8fea934: the four-module run passed 326 tests (4 skipped) and new-WSL passed 159. WorkflowInstallTests plus apply_claude_settings passed 40. The Windows Terminal and currency modules ran 64 tests whose only failure is the pre-existing auth_storage_failure test. validate.py, host receipts, catalogs, convergence, component matrix, grand list and ecosystem check each exited 0, SHA256SUMS returned 14 OK and both git diff --check forms exited 0.
• Not reproduced locally: the bootstrap modules, the 631-test 8-module total, the pre-push hook and gitleaks. At review time CI on 8fea934 showed native-token-tools, sota-sources and bootstrap-linux green, with the others still running.
• PR body: ### SOTA sources is present and complete, and the commit, row and count claims match the files. The 8 checked items hold, and the 2 unchecked ones (CI, review) were stated honestly.
• Tests: none was removed (apply_claude_settings went from 23 to 25, install_claude_profile from 79 to 94, native_token_ci from 47 to 48). The 3 changed tests are contract-assigned adaptations, and none was weakened.

Findings and dispositions:

1. should-fix, docs/harness-defaults.md ported rows: fixed in c1ae00392. Only the port's 17 rows changed, and main's rows stay byte-identical and in order. Four rows lost their whole-cell historical tag because their enforcer exists at the head: "Calling all function definitions methods", "Merging lifecycle hooks solely because their matchers match", "Checking only the first matching hook entry" and "Implicit Serena ancestor activation". The last one now names scripts/native_token_ci.py and the ported serena-no-project/ directory. Its receipts are the ported preimages, present at the head, so they carry their path rather than the tag. This is a deliberate reading of the fix's ":101 receipt mentions". In the ten rows with a current enforcer, every reference to the PR561-only CI-repair record now carries (historical: PR561 head fe4729d9) and links its fe4729d blob. That covers all four full ci-repair.json paths, and each cited section was confirmed in the blob first. The seven historical-only rows keep the tag closing their cell. The record now gives the split as ten current and seven historical-only rows and names them.
2. should-fix, missing discriminating controls in the body: fixed. This session reran both controls itself, on archive copies of 8fea9340 and of 3758ce52, and reproduced the reviewer's results. With the default steps including workflows, test_default_profile_never_reads_or_creates_workflows exits 1 with default run accessed workflows: <repo>/examples/claude-native/workflows/readiness-audit.js; as merged it exits 0. With the base 7d0174168 merger, the three hook-ownership tests exit 1 (failures=3: 2 != 4, 1 != 3, and a canonical-entry list diff); with the head merger they exit 0. Both pairs sit in a new "Discriminating controls" table.
3. minor, record base commit: fixed (wording). The record now names base 7d0174168. It says the port was prepared on cac8700ba and rebased onto that base, notes that docs/harness-defaults.md is byte-identical at both (empty diff), and links the 7d0174168 table.
4. minor, three blank lines before SecretGuardProfileTests: fixed (formatting only). One blank line was removed. The edit changes no behaviour and no assertion, the file is not one of the four byte-carried blobs, and it restores the fe4729d spacing.
5. minor, body evidence classes: fixed. The structure row and the merge row are relabelled structural_validation.
6. minor, record refresh link: fixed (wording). The quoted 2026-10-01T02:15:27Z figures now name native_lifetime_snapshot_20261001 in the CI-repair record as their source. The other link is labelled as the earlier 2026-09-30T22:52Z refresh, with its own 83,275,349 and 55,255,271 figures read from that blob.

No residuals.

New head: 3758ce5218321a421476248e97b70034a3c0ab2f. c1ae00392 holds the three content files. 3758ce521, the last commit, re-registers those three files in manifests/evidence.json; the generator checks passed, so no report was rewritten. It went out as a normal push. Main has moved to e0c329ae (#677); this branch was neither merged nor rebased.

Exit codes at 3758ce52 (nice 19, TMPDIR under /var/tmp, outside any sandbox):

• The contract's 8-module run exited 1: 631 tests, 36 skips and one failure, the pre-existing auth_storage_failure test, which fails the same way on a copy of e0c329ae. The four-module run exited 0 (326 tests, 4 skips), new-WSL 0 (159), WorkflowInstallTests 0 (15), apply_claude_settings 0 (25) and the docs-consistency module 0 (39 tests, 1 skip).
• validate.py exited 0, as did evidence_manifest.py --check, host_receipts.py validate, validate_catalogs.py, validate_convergence.py --all-recorded (29 valid), component_matrix.py --check, new_host_grand_list.py --check, build_ecosystem.py --check and SHA256SUMS (14 OK).
• git diff --check origin/main...HEAD exited 0 in both forms with no output. The contract's rebinding query exited 1 with empty output.
• The merge-tree landing check against e0c329ae exited 0 and printed LANDABLE. The pre-push hook exited 0.

Remaining before merge: no other-lane ack, since the custody contract requires none for lane:foundation and the new-WSL heads-up is informational. Still needed are the Codex root lane's exact-head read of 3758ce52 and 8/8 required checks on 3758ce52.

[seathatflowsinourveins]

Claude session native-agent-stack-5f: landing at head d5191821ef5ea6c0fa3ced17c855d7cf597de98e. The command center (wsl-architecture-design) gave its ACK at 2eff2970af7b, which reaches this head only through recorded mechanical refreshes whose owned non-registry patch-ids are equal (its carry rule); under the user's 2026-10-04 decision it owns merges that touch the hot registry. The read-only cross-family review (GPT-6.1 Sol max, the packaged Codex SDK worker) returned FINDINGS at 3758ce521832; this head is reached from it through recorded carry edges (equal owned patch-ids, or a cross-family delta read returning ACCEPT at the edge's target), with its 1 P1 item(s) resolved in one repair round as recorded.

Observed main f77a35eb2bf30bc4bf6f3b4ee51bc9ce5397b4c5. coordination/merge_tree_landing_check.py (sha256 a11649fcf2e5844c…) exit 0:

main f77a35eb2bf30bc4bf6f3b4ee51bc9ce5397b4c5 head d5191821ef5ea6c0fa3ced17c855d7cf597de98e base f77a35eb2bf30bc4bf6f3b4ee51bc9ce5397b4c5 merged-tree 9246e075d3b4138a588703ed62c1817bc1c2c282 merge-tree-exit 0
ok   1: clean three-way merge
ok   2: merged-vs-main paths 14, outside PR-owned 0 []
ok   3: main drift 0 paths, overlap with PR-owned inputs (registry excluded) []
ok   4: registry foreign rows equal True, order preserved True, PR-owned rows 13, unowned top-level keys differing []
ok   5: merged files[] sorted by path with no duplicates: True (9869 rows)
LANDABLE
condition 5: merged files[] sorted and unique (tree 9246e075d3b4138a588703ed62c1817bc1c2c282)

Required checks at this head: 8 pass . Unresolved review threads: 0.

[seathatflowsinourveins]

Claude session native-agent-stack-5f: post-merge observation. Landed as 33efbc3b6a46df556bdedf39cd8ea65762f38867 (parent f77a35eb2bf30bc4bf6f3b4ee51bc9ce5397b4c5), tree 9246e075d3b4138a588703ed62c1817bc1c2c282. The landed tree equals the checked merged tree and the parent is the observed main.