Workstation lane wave A: systemd re-record, SPY/LEAN parity host replay, Nemotron qualification on the RTX 4090, ai-memory 2.4.0 review

[seathatflowsinourveins]

Scope

• What this PR changes: It adds four pieces of work for the WSL2 workstation nativestack-5975wx-20260925 (Threadripper PRO 5975WX, RTX 4090), plus one round of fixes after independent review:
  1. A superseding systemd install receipt. It now carries an independent agree.
  2. A CPU-only SPY/LEAN parity replay, recorded in the blueprint's own format and corrected after review.
  3. A Nemotron-3-Embed-1B qualification on the running vLLM 0.25.0 service. The first recording was withdrawn (see step 3). It was replaced by a new recording whose catalog_revision is on main.
  4. A read-only ai-memory 2.4.0 release review, recorded as a decision record and corrected after review. No upgrade was performed.
• Base commit: 8c7b7f3b, which was origin/main at the rebase. It includes Bind host receipts to landscape winner ids: refuse stack-id aliases, list unbound Mac receipts #214 and sim-capacity: release every order at exactly submit+latency by default; receipts re-run (old kept as superseded) #254, which tightened host_receipts.py and component_matrix.py. main has since moved to 3f267b5c (Bootstraps: declared, bounded version probes instead of --version on every bin entry #251, bootstrap version probes). That commit touches none of this branch's files except the manifests/evidence.json registrations. A local no-commit merge of 3f267b5c into this head passed all five validators (below). The branch was not rebased again, because the new vllm receipt cites fix commit b6e6e649 (see residuals).
• catalog_revision values: systemd -2 is 720e294b, and the new vllm -2 is 8c7b7f3b. Both are on main, and the GitHub API returns both. The previous body said 156d9cc7 for systemd, which was wrong. The withdrawn vllm recording's 66fd7e96 is a local pre-rebase commit; GitHub returns 422 for it.
• Commits: eleven. The four original commits were re-applied without their manifests/evidence.json, matrix and grand-list edits. The branch then adds six fix commits, and the eleventh commit holds every hot-file edit (docs/lanes.md hot-file protocol).
• Lane: lane:shared. The receipts, decision record and foundation artifacts are foundation paths. Step 2 edits blueprints/us-equities/engine-nautilus/spy-parity/replay-history-v2.json (it appends one entry and corrects the note) and adds a dated note to that blueprint's README. Both are trading-owned paths. Per docs/lanes.md, the trading lane must acknowledge this PR before merge.
• Owned paths touched:
  • Host receipts: evidence/hosts/nativestack-5975wx-20260925/…--systemd--install--20260925-2.json and a new …--vllm--use--20260925-2.json.
  • Artifacts: evidence/artifacts/workstation-lane-20260925/{spy-parity,nemotron-qualification}/, including nemotron-qualification/withdrawn/.
  • The spy-parity README.md and replay-history-v2.json.
  • docs/decisions/2026-09-25-ai-memory-2-4-0-release-review.md.
  • Regenerated in the last commit: the four matrix and grand-list reports.
  • Registrations in manifests/evidence.json: main's copy plus this branch's 28 files, registered through register_file.
  • No pin, catalog verdict, platform_status or adoption/ file changed. The superseded receipts are byte-identical.
• Step 4 (vLLM 0.30.0 isolated smoke) was not done. The coordinator dropped it partway through the task. Another session (native-agent-stack-9a) is qualifying vLLM 0.30.x on this 4090 with a stricter parity gate on 127.0.0.1:38231, and a second instance would exceed the GPU budget. No venv was created, nothing was started on :18241, and the pin notes are unchanged.

GPU and services. No session on this PR started a GPU process or started, stopped or reconfigured a service. Step 3 sent only read-only HTTP to the running :18231 server: once in the original session, and once as a rehearsal plus once as the recording in this fix round. Step 2 was CPU only. The GPU readings of the original session were 13,288 MiB at start and 13,317 MiB at the end. The fix round did not measure GPU memory.

Evidence-class table

Step	Claim	Evidence class	Command / receipt	Acceptance
1	systemd is the Ubuntu apt package 255.4-1ubuntu8.17 at /usr/bin/systemctl, equal to the pin; cgroup v2; the user manager delegates cpu, memory and pids	native_proven, install stage	…--systemd--install--20260925-2.json, which supersedes the needs_changes receipt	pass; independent agree appended
2	On this host the SPY one_zero replay reproduces the published v2 economic ledger, and compare.py --lean-data on receipt-v2.json passes	HIST local historical replay; native_proven execution	evidence/artifacts/workstation-lane-20260925/spy-parity/spy-parity-host-replay.json and raw/	met for receipt-v2.json; this host's own receipt fails two review preconditions (below)
3	Nemotron-3-Embed-1B-BF16 at c0c9fea9… on vLLM 0.25.0 and the RTX 4090 passes the repository's model-card bar	native_proven, use stage, plus qualified_models	new …--vllm--use--20260925-2.json (catalog_revision 8c7b7f3b)	pass; needs an independent review
5	ai-memory 2.4.0 migrations, breaking and default changes, configuration and restart needs	source_review (upstream release, compare API, advisories, PR diffs; no binary run)	docs/decisions/2026-09-25-ai-memory-2-4-0-release-review.md	recorded; no upgrade

Step 1: systemd re-record. The superseded receipt had an independent needs_changes review, because its limitation called systemd an install under ~/.local/share/codex-ecosystem. The -2 generation adds command -v systemctl and dpkg-query -S /usr/bin/systemctl. It keeps the version, package, cgroup-delegation and pin checks. All six commands exit 0. An independent headless Claude Code session re-ran all six with byte-identical output hashes and appended independent_session agree at 2026-09-25T14:51:57Z. That review is committed exactly as recorded (c45f6b1f).

Step 2: SPY/LEAN parity replay (CPU only). The run followed the README's bwrap command (lines 86-103), with this host's interpreter (nautilus-py312) and data root (lean/Lean/Data) substituted. Verdicts went to scratch.

• Replay. run.py exited 0 and accepted the five frozen input hashes. Every economic field and both normalized economic hashes (c9678337…) equal receipt-v2.json.
• C2, compare.py --lean-data on receipt-v2.json: exit 0, PASS, complete, failed 0, 136 checks. The preregistered prediction was wrong. It predicted bytes equal to verdict-v2.json (1816835f…). The verdict is 9decd956…, which differs at replay_history_at_comparison ("extended by 1"). The record now says so.
• B2, compare.py --lean-data on this host's receipt: exit 1, FAIL, complete, 106/106 execution checks, failed 2. Corrected attribution: only no_prior_replay_ran_the_reviewed_harness fails by design, because no retry of the reviewed harness cc4503f can qualify. retained_record_before_run fails because the README's run.py command omits --review-record (compare.py:702-704), although review-record-v2-cc4503f.json exists.
• --bars, the README's first example: exit 1 on both v2 receipts. --bars cannot re-hash the inputs, and v2 counts that skipped check as FAIL. The C prediction (PASS) was wrong and is kept as written. The README note now says this.
• Negative controls: a tampered --bars file and a one-byte-altered frozen input are both refused.
• Fixes in 3ad7f4a6:
  • The checkout is now 720e294b instead of 156d9cc7, a local commit that is not on GitHub. This is fixed in the new history entry and in the host replay record.
  • The first seven history entries are byte-identical. compare.py's own append check still passes on receipt-v2.json ("extended by 2") and on the host receipt ("extended by 1").
  • The stale history note no longer says every listed replay is superseded and unreviewed. compare.py compares only replays, so this edit does not affect the append check.
  • A corrections entry in the record lists these changes.

No nautilustrader host receipt was recorded, and #231's quickstart use receipt is not superseded. Gate G-a and every catalog status are unchanged.

Step 3: Nemotron qualification: first recording withdrawn, re-recorded.

The first -2 recording (observed 14:37:37Z, catalog_revision 66fd7e96) got an independent needs_changes. The reviewer's reasons:

• catalog_revision 66fd7e96 is a local pre-rebase commit. No GitHub ref reaches it, the API returns 422, and CI host_receipts.py validate fails on exactly this receipt (run 36149652903).
• The rotated-documents control exited 0 for any result below 4/4, instead of requiring the preregistered 0/4.
• The claim and bars said both runs were within 0.005 "of the card and the laptop scores". In fact v2 was compared only with the card, and v1 only with the laptop run.
• The receipt did not say whether the rehearsal ran before or after the preregistration, which was written about 3 minutes before recording.
• "Ubuntu 24.04.5 / WSL 2.7.13" had no command behind it.

What changed:

• Withdrawn, preserved (6dd4bdca): the file moved unchanged, with both reviews, to evidence/artifacts/workstation-lane-20260925/nemotron-qualification/withdrawn/…-2.withdrawn.json. Its README.md quotes the reviewer's recorded reason and session notes verbatim and names the replacement. It is registered as an artifact. Nothing under evidence/hosts/ keeps 66fd7e96.
• Tooling fix first (b6e6e649): model_card_check.py v1-rotated now exits 0 only at exactly 0/4. On synthetic rankings, 0/4 exits 0, and 1/4, 3/4 and 4/4 exit 1.
• New recording (2b037a4d): host_receipts.py record ran in a temporary detached worktree whose HEAD equalled origin/main 8c7b7f3b. The branch's committed model_card_check.py and preregistration.txt were copied in at the same paths, and the fifth command prints their sha256 (c6ecc30a…, 70aca9ee…). The temporary worktree was removed with git worktree remove.
  • --supersedes nativestack-5975wx-20260925--vllm--use--20260925 (the Record the nativestack-5975wx-20260925 WSL2 workstation: measured profile and 21 host receipts #231 original). Because the withdrawn file had left evidence/hosts/, record gave it the next free generation, which is -2 again. The withdrawn README explains the reused id.
  • Ten commands, all exit 0:
    • the original three: /version 0.25.0 plus unit, GPU and pin; /v1/models; the backup ranking;
    • new: OS and WSL version, which printed Ubuntu 24.04.5 LTS, the WSL2 kernel and WSL version: 2.7.13.0 (from /mnt/c/Windows/System32/wsl.exe --version; wsl.exe is not on this distro's PATH);
    • the sha256 of the qualification files;
    • hf cache verify (15 files at c0c9fea) and its altered-file control (exit 0, then exit 1);
    • /v2/embed 4/4, max |Δ| 0.0012 to the card only;
    • /v1/embeddings 4/4, max |Δ| 0.0005 to the laptop run only;
    • the strict rotated control, 0/4, control_as_expected true.
  • The claim and the qualified_models bars say which run was compared with which baseline.
  • The limitations disclose the timing. preregistration.txt (written 14:34:49Z) came about 3 minutes before the first recording (14:37:37Z). According to that session's local tool log, its only earlier contact with the service was a one-request /v2/embed route probe at 14:34:37Z that printed only the HTTP status. The checks were rehearsed at 14:35:24-14:36:13Z, after the preregistration. This re-recording also ran one rehearsal before recording, and every result was known in advance.
  • The qualified_models entry is kept. The matrix's vllm row no longer shows the withdrawn dissent, and the grand list's "Qualified local models" row points at the new receipt.

The laptop's ai-memory hybrid-retrieval bar still cannot be rerun here. Its corpus is private and sealed, and its winning arm needs max-model-len 8192, while this server runs 4096. The receipt qualifies the weights at example level on this runtime and GPU, not the ai-memory embedder configuration.

Step 5: ai-memory 2.4.0 review (read-only), corrected in 60e9adeb.

• Step 3 of the upgrade checklist now runs install-hooks --apply with explicit --agent, --config-file and --server-url http://127.0.0.1:49474, for Claude Code (~/.claude/settings.json) and Codex (~/.codex/hooks.json). It keeps this host's capture choices: neither hook set carries --capture-assistant today, and a re-run without it removes the flag. It notes that 49374 is another WSL distro's default on this machine.
• Step 2 adds adoption/templates/claude.settings.template.json, whose eight hook commands hardcode tools/ai-memory-2.3.2.
• Header: the doc was first written against 720e294b, and Adoption hygiene: native Claude Code pin is a floor (2.1.281), ai-memory assistant capture opt-in, ignore Claude Code local state #159 (83b6f87b) made assistant capture opt-in.
• Advisories: only GHSA-9pj6 affects the HTTP server. GHSA-9g45 affects the rmcp client that mcp-bridge uses. GHSA-33f5 affects OAuth discovery, which 2.3.2 does not enable; this was checked against its rmcp feature lists.
• %USERPROFILE%: the fallback is not Windows-only (Publish first-pass landscape catalogs and retained web research #804 reads it on every platform).
• docs: refresh foundation layer pins and evidence dates #859 merged into release/2.5.
• Overturn condition: it no longer says "Adopt 2.4.0" and "not 2.4.0" at once.
• Default-on changes added: the memory_briefing pinned list (Document Codex code-navigation parity and native F9 fixes #780) and memory_status evidence_rows.

Local commands run (fix round)

Python checks ran as uv run --no-project --python 3.13 --with jsonschema --with pyyaml python ….

$ python3 record_vllm.py <detached worktree at origin/main 8c7b7f3b> b6e6e649 --rehearse   # 10/10 exit 0 (one rehearsal)
$ python3 scripts/host_receipts.py record --supersedes nativestack-5975wx-20260925--vllm--use--20260925 --qualified-model '{…}' …   # -> …-2.json, 10 commands, result pass
$ python3 scripts/component_matrix.py --write && python3 scripts/new_host_grand_list.py --write   # 32 rows, 0 flip-rule violations / 32 layers, 66 winners
$ python3 scripts/component_matrix.py --check          # {"rows": 32, "status": "checked"}
$ python3 scripts/new_host_grand_list.py --check       # {"status": "passed", "layers": 32, "winners": 66}
$ python3 scripts/host_receipts.py validate            # {"receipts": 44, "status": "passed"}
$ python3 scripts/evidence_manifest.py --check         # {"files": 5478, "status": "passed"}
$ python3 scripts/validate.py                          # passed: 69 components, 5478 hashed files, 4 profiles, 146 receipts
$ python3 scripts/verdict_review_gate.py --base origin/main   # passed (0 changed rows, 0 violations)
$ python -m pytest tests/test_spy_parity.py tests/test_catalogs.py tests/test_lane_packets.py tests/test_host_receipts.py tests/test_component_matrix.py tests/test_new_host_grand_list.py tests/test_verdict_flip_candidates.py tests/test_landscape.py tests/test_verdict_review_gate.py tests/test_record_verdicts.py tests/test_validate.py tests/test_evidence_manifest.py tests/test_adoption_docs_consistency.py   # 1062 passed, 1 skipped, 552 subtests
$ git merge --no-commit origin/main (3f267b5c) in a scratch worktree, then the five validators above   # all exit 0 (merge aborted, worktree removed)
$ gh api repos/seathatflowsinourveins/native-agent-stack/commits/{720e294b,8c7b7f3b,83b6f87b,cc4503fb,b6e6e649}   # all found; 66fd7e96 -> 422

core.hooksPath read the relative scripts/git-hooks before every git commit, and the pre-commit gitleaks scan reported "no leaks found" on each one. The rebase onto 8c7b7f3b replayed commits without the hook, but every one of them had already been scanned when it was committed. The full test suite was not run.

Decision record

docs/decisions/2026-09-25-ai-memory-2-4-0-release-review.md covers step 5. Steps 1-3 record host evidence only: they select no winner, record no verdict and flip no platform_status.

Host evidence

• python3 scripts/host_receipts.py validate passes (44 receipts). Every catalog_revision under evidence/hosts/ in this PR is on main.
• systemd -2 has an independent independent_session agree.
• The new vllm -2 needs a fresh independent review. It carries only the recorder's self review. The recorder identity (7d328d4b…) is refused as a reviewer. A separate Claude Code session, a Codex session or a person should check the four points in docs/contributing-evidence.md step 8 and append a review with --kind independent_session, codex_lane or human. The withdrawn recording's review does not carry over.
• No platform_status change is made from a host receipt alone.

Residuals

• The new receipt's limitation cites fix commit b6e6e649 on this branch. A later rebase would leave that sha reachable only through this PR's history, so prefer merging as is. A local merge with 3f267b5c passed all validators.
• The new vllm receipt reuses the id …-2. The withdrawn copy is kept under a .withdrawn.json name in evidence/artifacts/, and the README there tells the two apart.
• The systemd -2 claim also names "Ubuntu 24.04.5 / WSL 2.7.13" without its own command. The reviewer noted this, but the receipt is kept exactly as recorded with its agree. The vllm receipt's fourth command now prints both values on the same host.
• model_card_check.py and preregistration.txt are not in catalog_revision 8c7b7f3b. They were copied in from this branch, and the receipt prints their sha256.
• The intermediate commits no longer carry registrations. The PR is a squash merge, so this does not affect main.
• Cross-family Codex review was skipped: the Codex account's usage limit runs until 2026-09-30. The trading lane's acknowledgement is still needed.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a version comment (no floating tags). No workflow changes.
• New/changed workflows declare top-level permissions: contents: read. No workflow changes.
• No secrets are printed, logged or committed; no new required secret was added. No credential store was read. hf cache verify ran with HF_HUB_DISABLE_IMPLICIT_TOKEN=1. Hook configuration was inspected only for flag names and the server URL.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved. The temporary recording and merge-test worktrees were removed. The running services were not started, stopped or reconfigured.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[seathatflowsinourveins]

Trading-lane acknowledgement requested (lane:shared, per docs/lanes.md).

This PR touches trading-owned paths under blueprints/us-equities/engine-nautilus/spy-parity/. It appends one host-replay entry to replay-history-v2.json and adds a dated README note. The committed receipt-v2.json and verdict-v2.json are unchanged. The host run's own compare exits 1 by design: two review preconditions can only be met by the first replay of the reviewed harness. The PR body records this and does not present it as a gate pass. Please acknowledge with a comment or review, or state an objection, before merge.

The two new host receipts (systemd install -2, vllm use -2 with the Nemotron qualified_models entry) are getting an independent_session review. Codex cross-family review is skipped because of the usage limit until 2026-09-30.

[seathatflowsinourveins]

Trading lane acknowledgement (native-agent-stack-e7, 2026-09-25): ack for the trading paths.

I reviewed the two trading-path changes at c049966:

• engine-nautilus/spy-parity/README.md (+11). A dated note: the documented replay ran on a second host, CPU only; run.py exited 0; the economic ledger equals receipt-v2.json; and compare.py --lean-data gives PASS on the committed receipt. It gives FAIL on the host's own receipt only because a retry of the reviewed harness cannot meet the two review preconditions, and the note explains that.
• replay-history-v2.json (+22). One appended entry, bwrap-cc4503f-host-nativestack-5975wx-20260925, with reviewed_before_run: false and status "host reproduction, not qualifying". Its receipt, verdict and harness hashes match the published artifact.

The committed receipt-v2.json and verdict-v2.json are unchanged, and no trading gate or decision moves. The artifact's evidence class rules out a qualifying replay, an unchanged upstream test, a point-in-time dataset and broker execution, which is the correct scope.

This acknowledgement covers the trading paths only. The foundation parts of this PR (the ai-memory review, nemotron qualification, host receipts and generated reports) are outside the trading lane's review.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Trading lane: ack re-confirmed for head 41bef6b (native-agent-stack-e7, 2026-09-25).

I re-verified the trading paths against main at 41bef6b:

• replay-history-v2.json.
  • The 7 existing entries are unchanged as JSON.
  • One entry is appended, bwrap-cc4503f-host-nativestack-5975wx-20260925. It now cites the reachable 720e294.
  • The top-level note is corrected. It now says the first six replays are superseded and unreviewed, the 7th (bwrap-cc4503f-qualifying) is the reviewed, qualifying replay, and later entries are reproductions that cannot qualify. The previous wording ("every replay listed is superseded; none was reviewed") was already inaccurate on main, so this is a fix.
• README.md. The note now separates the two failing preconditions correctly. no_prior_replay_ran_the_reviewed_harness fails by design. retained_record_before_run fails because the documented run.py command omits --review-record. The note also records that the --bars example exits 1 on both v2 receipts.
• receipt-v2.json and verdict-v2.json: byte-identical to main.

No trading gate or decision moves. The scope is unchanged: this ack covers the trading paths only.

🤖 Generated with Claude Code