Record the nativestack-5975wx-20260925 WSL2 workstation: measured profile and 21 host receipts

[seathatflowsinourveins]

Scope

• What this PR changes:
  • The measured hardware profile of a new WSL2 workstation, host id nativestack-5975wx-20260925, written by scripts/hardware_profile.py --record-host: evidence/artifacts/hw-profiles/nativestack-5975wx-20260925/profile.json and a new native_proven entry in adoption/hardware-profiles.json hosts[]. The labelled projection wsl-workstation-128gb-projected stays, as docs/next-host-stages.md describes. The grand list now shows the measured host next to it.
  • 21 host receipts under evidence/hosts/nativestack-5975wx-20260925/: 14 use-stage native_proven, 1 use-stage local_integration, and 6 install-stage native_proven. All 21 pass. None has an independent review yet (see Host evidence).
  • The regenerated catalogs/landscape/component-evidence-matrix.json, docs/component-evidence-matrix.md, catalogs/landscape/new-host-grand-list.json and docs/new-host-grand-list.md, plus the manifests/evidence.json registrations.
• Base commit: aa0812b. Every receipt was recorded from a worktree at aa0812b1 (then origin/main), so catalog_revision stays reachable after a squash merge.
• Owned paths touched: the files above only. adoption/pins-*.json, adoption/platforms/linux-wsl2.md, docs/secret-storage.md, docs/acceptance-evidence-policy.md, docs/next-host-stages.md and docs/new-workstation-runtime-profile-20260922.md are untouched.

The host, as measured. Ubuntu 24.04.5 LTS in WSL 2.7.13.0 (kernel 6.18.33.2-microsoft-standard-WSL2), distro NativeStack. AMD Ryzen Threadripper PRO 5975WX with 48 visible CPUs (.wslconfig processors=48). 102.2 GiB in /proc/meminfo (.wslconfig memory=104GB, swap=24GB). NVIDIA GeForce RTX 4090 24 GB (driver 616.64). systemd 255 (255.4-1ubuntu8.17) on cgroup v2; the user manager delegates cpu, memory and pids.

• Tiers: large-32b-q4 generation, headroom semantic RAG, concurrency cap 16, bounded-run memory 25.6 / 17.1 GB. These are the projection's tiers.
• .wslconfig differs from the decided profile. docs/next-host-stages.md records memory=112GB, processors=60, swap=16GB; this host has 104GB, 48 and 24GB. That page is owned by another PR, so it is not edited here.
• Second physical machine. Every receipt declares second_physical_machine: true. The catalog's WSL authoring host is an Intel Core Ultra 9 275HX laptop with an RTX 5090 Laptop GPU (evidence/artifacts/sota-refresh-20260923/hw-profiles/this-host.json), and this host's measured CPU and GPU differ from it. The claim is self-declared; the reviewer should confirm it.

Evidence-class table

Paths are evidence/hosts/nativestack-5975wx-20260925/nativestack-5975wx-20260925--<component>--<stage>--20260925.json.

Receipt	Stage	Evidence class	What the retained commands show
gitleaks	use	native_proven	gitleaks resolves to the guarded launcher, whose hashes and its runner's equal adoption/tools. The checkout's latest commit and a throwaway clean commit: no leaks. Positive control: the gitleaks sidekiq-secret fixture commit exits 1 with exactly one finding (RuleID: sidekiq-secret, secret redacted).
rtk	use	native_proven	rtk git status is compacted (599 to 228 bytes). rtk proxy git status is byte-identical to git status. A missing ref returns git's own exit 128.
qmd	use	native_proven	BM25 search on this host's nativestack-catalog index (15 files, 0 vectors): the expected top hit adoption/tools/README.md, and [] for an absent token.
prometheus	use	native_proven	The running 3.14.0 server: up for 7 scrape targets (all 1), 240 vLLM samples in the last hour, 2 rule groups and 14 rules with health ok. Five firing alerts are named (see Findings).
loki	use	native_proven	The running 3.7.8 server: stored service_name values and 131,557 lines over 24 h. A line pushed with the native push API (HTTP 204) comes back from a LogQL query.
opentelemetry-collector-contrib	use	native_proven	The running 0.161.0 Collector took one OTLP/HTTP log into Loki. The body came back as [content omitted], the allowlisted observation_id was kept, and both non-allowlisted attributes were dropped.
vllm	use	native_proven	The running server reports /version 0.25.0 on the RTX 4090. Nemotron-3-Embed-1B returns 2048-dimensional embeddings, and the backup passage ranks first (0.5844 against 0.2269 and 0.0588).
qdrant	use	native_proven	The running 1.19.1 server: a temporary collection of real embeddings is created, upserted and searched (backup passage first), then deleted. It returns 404 after deletion, and the collection count is 5 before and after.
restic	use	native_proven	A throwaway repository with a run-time password: adoption/ (51 files) is backed up, check --read-data is clean, and restore --verify plus diff -r give an identical tree. Positive control: after one pack is corrupted, check exits 1.
dagu	use	native_proven	Scratch DAGU_HOME: a two-step DAG that passes output between steps succeeds. Negative case: a step exiting 3 gives Result: Failed and error: exit status 3.
difftastic	use	native_proven	Real revisions 57fb6d89 to aa0812b1: syntactic change, exit 1. Control: a whitespace-only JSON reformat, where cmp differs and difft exits 0.
worktrunk	use	native_proven	Throwaway repository, no user or project config, stdin /dev/null: switch --create, list, remove --foreground, then only main remains.
markitdown	use	native_proven	A real 8-K's HTML (the vendored edgartools fixture, sha256 a1e7ca47…) becomes 5,536 bytes of Markdown with 29 table rows and no raw tags.
nautilustrader	use	native_proven	The unchanged upstream quickstart at 1b0a49d2 (sha256 verified) runs in Bubblewrap with loopback only and no /home: 10,000 iterations, 1,804 events, 902 fills, 451 positions, 903 account rows, ending total 1000431.00. account.csv is byte-identical to the authoring host's accepted runs.
poppler	use	local_integration	The frozen document-ingestion protocol on the repository's designed corpus: all 8 checks are true, including the native invalid-PDF refusal.
codex	install	native_proven	0.155.1: path, bootstrap and catalog pin, --version. Readiness only.
ai-memory	install	native_proven	2.3.2: path, pins, --version. Readiness only.
mcporter	install	native_proven	0.13.13: path, pins, --version. Readiness only.
systemd	install	native_proven	255.4-1ubuntu8.17 by systemctl and dpkg-query, cgroup v2, cpu/memory/pids delegated. Readiness only.
headroom	install	native_proven	0.37.0: path, catalog pin, --version. Readiness only.
ccusage	install	native_proven	20.0.24: path, catalog pin, --version. Readiness only.

The install receipts are version and pin checks. They show the pinned build is present and starts, which is readiness evidence, not functional acceptance, and they support conditional at most. The use receipts against running services observed services that were already up under their systemd user units; nothing was started, stopped or reconfigured.

Rehearsal and one post-recording edit. Each receipt's commands were rehearsed once through the recorder's own execution path (/bin/sh from the repository root, sanitized 400-character excerpt) before recording. The rehearsal found and fixed these problems:

• a false pass in restic's positive control: restic writes packs read-only, the overwrite failed, and the stale exit status satisfied the check. The recorded command now makes the pack writable, exits 2 if the overwrite fails, and requires an error in the check log;
• restic 0.19.1's restore layout;
• gitleaks printed no finding lines without --verbose;
• a /home skeleton inside the Nautilus sandbox, because the environment lives under $HOME. It is now bound at /opt/nautilus-env, which the receipt states as its one difference from the accepted argv;
• a scratch directory named home, which the sanitizer redacted;
• pin, path and collection lines that some claims cited but no command printed.

After recording, the restic receipt's claim was reworded to match its retained excerpt: pack … contains 2 errors and Fatal: repository contains errors, not the rehearsal's ciphertext verification failed. Its commands, outputs and timestamps are unchanged, and its hash was re-registered. Between them, the rehearsals and recordings wrote four labelled synthetic probe records to this host's Loki, which has 72-hour retention. Two went through Loki's push API. Two went through the Collector, which also appended them, with the body already replaced, to its events file. They also created two temporary Qdrant collections, both deleted.

Not recorded, and why

• claude-code: installed 2.1.282, at or above the 2.1.280 bootstrap floor, but the landscape winner pin is still 2.1.278. A receipt would need --allow-unbound-version and would never count. No claude -p model run was attempted.
• candidate:cli-cli (gh 2.101.0) and candidate:astral-sh-uv (uv 0.12.17): both winners are unpinned, so no receipt can bind. node 24.21.0 is not a catalog component.
• codex use: no codex exec model turn was attempted. The account's Codex usage limit was reported to run until 2026-09-30; this PR did not verify that.
• Use stage for ai-memory, mcporter, headroom, ccusage and systemd: not recorded. The memory store and usage logs are private, and the mcporter config is machine-specific client configuration. headroom needs its proxy or MCP server. systemd use would start transient units, which this task excluded.
• IBKR environments (nautilus-ibapi-10.45.1-py312, nautilus-ib-1.231.0-py312): IBKR local acceptance is a later gate (docs/next-host-stages.md step 6). The skills and agents installed today (typesafe-ai, gh-fix-ci, security-best-practices; 7 catalog agents) either have unpinned winners or need a model-mediated invocation.
• Model qualification on the RTX 4090: not run. The vllm receipt records serving only, with no qualified_models.
• Not recorded in this PR: socraticode, serena, context-mode and jcodemunch (MCP servers used through the clients); syft, grype and zizmor; hf; Grafana, Alertmanager and ntfy (not landscape winners); the llama.cpp generation server (a b11146 CUDA build; the stack's llama-cpp entry is b11057). For sandbox-runtime, srt --version prints 1.0.0 while its package is 0.0.77, so a version call cannot back that pin.
• Not installed on this host: promptfoo, playwright-test (no pnpm), agent-browser, openresearch, tavily-cli, and the pandera, mlflow, dvc and inspect-ai rows.

Findings (left unchanged here)

1. EcosystemCollectorHealthUnexpected fires on this host. The rule, rendered from observability/backends/templates/ecosystem-prometheus-rules.yml.example, probes 127.0.0.1:14333, but this host's Collector health_check listens on, and its http_check probes, 127.0.0.1:24333. The other four firing alerts are for services this host does not run: OmniRoute, FreeLLMAPI, the Dagu dashboard and paper-trial metrics.
2. The claude-code winner pin lags the bootstrap floor: landscape 2.1.278 against floor 2.1.280 and 2.1.282 installed here.
3. This host's QMD index is nativestack-catalog, not the native-agent-stack-catalog that AGENTS.md names. It indexes only the adoption collection and has no embeddings yet.
4. verdict_flip_candidates.py lists 2 new linux-wsl2 candidates, difftastic and worktrunk in foundation/git-github-automation: declared not_established, and the evidence now supports conditional. Report only; a flip needs the lane re-record.
5. The no-CPU refusal test from Stop the no-CPU test slice so guarded-runner runs stop leaking systemd units #229 fails on this host. In the dropped subtest, the scope created in the fresh ecosystem-check-nocpu-<pid>.slice has no memory.max, so ecosystem-bounded-run refuses with memory.max is not enforced where the test asserts cpu.max is not enforced. The refusal itself holds (exit 78, command not started); only the asserted reason differs. This session's two runs stopped their slices, so Stop the no-CPU test slice so guarded-runner runs stop leaking systemd units #229's cleanup works here. Four older ecosystem-check-nocpu-<pid>.slice units from earlier runs are still loaded on this host and are left in place.
6. The pre-commit gitleaks hook does not run on this host. The repository-local core.hooksPath is an absolute path to scripts/git-hooks in the main checkout, and that directory does not exist there. Git skips a missing hooks directory silently, so commits from any worktree get no staged-change scan. scripts/git-hooks/pre-commit documents the relative form git config core.hooksPath scripts/git-hooks, which resolves inside each checkout. This commit's range was scanned manually instead. The host configuration is left unchanged.

Local commands run

Each command ran as uv run --no-project --python 3.13 --with jsonschema --with pyyaml python …, because the host's python3 lacks jsonschema.

$ python3 scripts/hardware_profile.py --record-host nativestack-5975wx-20260925 --label "WSL2 x86_64 workstation, …"
{"evidence": "evidence/artifacts/hw-profiles/nativestack-5975wx-20260925/profile.json", "host_id": "nativestack-5975wx-20260925", "status": "recorded"}
$ python3 scripts/host_receipts.py record …          # 21 receipts, each result=pass
$ gitleaks dir evidence/hosts/nativestack-5975wx-20260925 --no-banner --no-color --redact            # no leaks found
$ gitleaks dir evidence/artifacts/hw-profiles/nativestack-5975wx-20260925 --no-banner --no-color --redact   # no leaks found
$ gitleaks git --no-banner --no-color --redact --config .gitleaks.toml --log-opts="origin/main..HEAD" .   # 1 commit, ~144.6 KB, no leaks found (the pre-commit hook did not run; see Findings 6)
$ python3 scripts/component_matrix.py --write          # {"flip_rule_violations": 0, "rows": 32, "status": "written"}
$ python3 scripts/new_host_grand_list.py --write       # {"status": "written", "layers": 32, "winners": 66}
$ python3 scripts/host_receipts.py validate            # {"receipts": 37, "status": "passed"}
$ python3 scripts/evidence_manifest.py --write && python3 scripts/evidence_manifest.py --check   # 5360 files: written, passed
$ python3 scripts/validate.py                          # passed: 69 components, 5360 hashed files, 4 profiles, 145 receipts
$ python3 scripts/verdict_review_gate.py --base origin/main   # passed: 2 regenerated catalogs/landscape files, 0 changed rows, 0 violations
$ python3 scripts/receipt_staleness.py                 # 34 buckets, 0 flagged
$ python3 scripts/component_matrix.py --check          # 32 rows, checked
$ python3 scripts/new_host_grand_list.py --check       # passed
$ python3 scripts/validate_catalogs.py && python3 scripts/validate_foundation.py --root . --json && python3 scripts/landscape.py --root .   # passed, errors []
$ python3 scripts/build_ecosystem.py --check           # passed
$ python3 scripts/verdict_flip_candidates.py           # 8 candidates (6 pre-existing macOS, 2 new linux-wsl2)
$ python3 scripts/release_due.py --strict-if-repinned origin/main   # release_due, non-strict, exit 0: pre-existing tag drift; this PR's 4 generated paths already differed from v2026.09.24.1
$ python3 -m unittest                                  # Ran 4720 tests in 395.5s: FAILED (failures=1, skipped=534)
$ python3 -m unittest tests.test_adoption_guarded_runners.BoundedRunCpuQuotaTests   # Ran 7 tests: FAILED (failures=1), same test

The one failure is tests.test_adoption_guarded_runners.BoundedRunCpuQuotaTests.test_scope_with_a_dropped_or_altered_cpu_quota_is_refused_with_78, in its launcher='dropped' subtest. It reproduces in isolation, and this PR does not cause it: tests/, adoption/tools/ and scripts/ are byte-identical to origin/main. See Findings 5.

Decision record

None. This PR records host evidence only: it selects no winner, records no verdict, and flips no platform_status.

Host evidence

• python3 scripts/host_receipts.py validate passes for every new/changed receipt (37 receipts, passed).
• Independent review is recorded (scripts/host_receipts.py review) or explicitly requested in this PR. Requested: each receipt carries only the recorder's self review. A reviewer in a separate session should check the four points in docs/contributing-evidence.md step 8 (layer role, positive control, backed claims, bound to the winner) and append a review. The reviewer must be a different Claude Code session, a Codex session or a person; this recording session's identity is refused. Cross-family Codex review was skipped: the Codex account's usage limit runs until 2026-09-30. With an agree review, the 14 use-stage native_proven receipts become qualifying receipts for their winners on linux-wsl2-x86_64.
• No platform_status change is made from a host receipt alone. adoption/manifest.json and catalogs/landscape/{foundation,us-equities}.json are unchanged.

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. The receipts were re-read and scanned for home paths, usernames, session ids, emails and token patterns (none found), and gitleaks found no leaks. No credential store was read. The restic probe's run-time password was deleted with its throwaway repository.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved.

Independent review (appended 2026-09-25)

A separate headless Claude Code session reviewed all 21 receipts against docs/contributing-evidence.md step 8. It had its own session id, so it was not the recorder. It re-ran at least one command per receipt on this host. The verdicts are recorded with host_receipts.py review --kind independent_session:

• 20 agree. Use receipts reproduced live include:
  • restic backup, restore and the corruption control (exit 1),
  • vLLM embedding ranking (cosines within 0.0004),
  • Prometheus: 7 targets up, 14 healthy rules,
  • the full NautilusTrader quickstart with a byte-identical account.csv,
  • qmd, rtk, worktrunk, dagu, difftastic and gitleaks with their controls.
• 1 needs_changes: systemd (install). It fails point 3. Its generic limitation says the tool is an install under ~/.local/share/codex-ecosystem, but systemd is the Ubuntu package at /usr/bin/systemctl. Its version and pin claims were reproduced. The receipt stays withheld with its veto intact.

A same-day re-record would have silently overwritten this receipt and its review. That was caught before commit and reverted. The tool defect is fixed in a separate PR: record now refuses to overwrite and supports an explicit supersede.

Rebased onto d993d5ff (#230). Regenerated component_matrix.py --write (32 rows, 0 flip-rule violations) and new_host_grand_list.py --write (32 layers, 66 winners). host_receipts.py validate passes (37 receipts), as do evidence_manifest.py --check (5366 files), validate.py and verdict_review_gate.py (0 changed rows). Codex cross-family review is still skipped because of the usage limit until 2026-09-30.

Residuals:

• The systemd install receipt is withheld and can be re-recorded with the corrected limitation once the supersede support lands.
• claude-code, gh and uv have no bound pins, so they got no receipts.
• IBKR environments, skills and agents are not recorded here.

🤖 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.