Preserve #410 Mac model-hosting observations with scoped corrections

[seathatflowsinourveins]

Scope

• Preserve Mac-native memory/RAG model hosting on mac-coordinator-64gb-20260925 (#379): measured capacity, receipt, host-roles proposal #410's Mac session measurements from 2026-09-27 as descriptive reference evidence, with the eight measurement/reducer files byte-identical to 14b5c52145d9cc42d4c9da36df53ac994009856a. Correct the README, add a provenance/scope record, date two macOS-page passages and port the verified unpublished-revision anti-pattern append.
• Base commit: 463a57b983eec540ae90eb45c2b1a7c6fc469aed (main). Head: bef836c62b25af872db5261be94c46080ec25f17, after one review repair round on e65c612e5c3a69be1b95957341bed48c3a663d85. The content was prepared on cac8700ba914950266272347468bff7ad630a4bf and rebased onto this main without conflict. Main's drift since then touched no port path except manifests/evidence.json, which was re-taken from main and re-registered in e65c612e, that round's registry commit. The repair round's last commit, bef836c62, re-registers only the two corrected files on the branch's own registry.
• Main has since moved to 6112d14d40f741855f0b961124d98939daaad00e (GPT-6 lane: opt-in native-to-OmniRoute failover on a native usage limit #649, which also changed manifests/evidence.json). The repair round did not rebase. The pre-merge rebase under the hot-file protocol remains a merge gate: take main's registry, re-register the 12 port paths, run both --write and both --check commands, then update this body's base, head and counts.
• Lane: lane:foundation.
• Owned paths touched: evidence/artifacts/mac-model-hosting-20260927/, adoption/platforms/macos-arm64.md, the existing unpublished-revision row in docs/harness-defaults.md, and file registrations in manifests/evidence.json.

SOTA sources

• ggml-org/llama.cpp b11057, 59657a613ab0fa4ab327d6c790123dff30bfbd67, common/arg.cpp L4235 registers --spec-type; common/speculative.cpp L37 registers draft-mtp. Both exact files and the tag/release identity were verified.
• unsloth/Qwen3.8-27B-GGUF at 4ca720788d1e01f1bff70c033e0d0028fd02e502: publisher revision and Qwen3.8-27B-UD-Q4_K_M.gguf filename verified.
• ggml-org/embeddinggemma-300M-GGUF at 0f741b5a6585bd53aeb15cd1372c56f2a0f65e12: publisher revision and embeddinggemma-300M-Q8_0.gguf filename verified.
• nvidia/Nemotron-3-Embed-1B-BF16 at c0c9fea93ea424587517f2c59e20db9f1d6bf615, config.json: Ministral3Model, is_causal: false, hidden size 2048 verified.
• GitHub: managing automatic branch deletion, with the repository's returned delete_branch_on_merge: true setting and the failed first #391 validate job. Its later attempt 2 passed.
• actions/checkout at 3d3c42e5aac5ba805825da76410c181273ba90b1, README full-history example and src/ref-helper.ts L69: branch/tag fetching verified against the pin used by the four full-history checkouts in validate.yml.

Evidence-class table

Classes follow docs/acceptance-evidence-policy.md; the artifacts retain the recording session's embedded labels byte-identically.

Claim	Evidence class	Command / receipt
llama-bench throughput on the Mac on 2026-09-27	Upstream example or native operation	llama-bench -p 512,4096 -n 128 -ngl 99 -fa on -r 3; historical capacity.json
Embedding dimensions/cosine observation	Local integration check	tools/adoption/embed_acceptance.py; historical embed-acceptance.json; no discriminating control, so no acceptance
li26 reported quality/equality summary	Local integration check	li26_mac_descriptive.py; historical aggregates; per-filing rows and private-run hashes unavailable
Serving (with its host memory, swap and jetsam readings), co-residency and gate timings	Local integration check	Historical capacity.json, coresidency.json, gate.json; co-residency build unresolved and gate conditions incomplete
Byte identity of all eight measurement/reducer files	Independent observation	git diff --exit-code from 14b5c521 to this head over the eight files exits 0 with no output; the builder's earlier blob-to-file cmp of each file also exited 0
Manifest integrity, generated reports and repository contract tests	Structural validation	Native repository validator/check commands and the requested unittest modules, with results below

Local commands run

Repair round: every command below ran on the committed head bef836c62b25af872db5261be94c46080ec25f17 at niceness 19, with TMPDIR set to a scratch directory outside every repository. Standard output and standard error went to files, and each exit code is the command's own, never a pipe's. No environment override was needed.

$ python3 scripts/validate.py
exit 0: 69 components, 9572 hashed files, 4 profiles, 193 receipts, passed.
Integrity and scope checks only; no live provider or GPU execution.

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

$ 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.
Both checks passed after the re-registration, so no --write ran in this round.
$ python3 scripts/evidence_manifest.py --check
exit 0: 9572 files, passed.

$ python3 -m unittest tests.test_adoption_bootstrap_macos tests.test_adoption_docs_consistency tests.test_adoption_launchd tests.test_ecosystem_manifest tests.test_host_requests
exit 0: Ran 399 tests in 97.114s, OK (skipped=35).

$ git diff --check origin/main...HEAD
exit 0, no output.
$ git diff --exit-code 14b5c52145d9cc42d4c9da36df53ac994009856a HEAD -- evidence/artifacts/mac-model-hosting-20260927/{capacity,coresidency,embed-acceptance,gate,li26-mac-descriptive,li26-window-samples,swap-event}.json evidence/artifacts/mac-model-hosting-20260927/li26_mac_descriptive.py
exit 0, no output: the eight files are byte-identical to the #410 head.
$ git diff --name-only origin/main...HEAD
exit 0: 13 paths, the 12 content files and manifests/evidence.json.
The contract's forbidden-path grep -E over that listing exits 1: no hits.
$ grep -n -i -E 'qualif|native_proven' evidence/artifacts/mac-model-hosting-20260927/README.md
exit 0; the same five scoped/negated hits, listed below.

$ git push origin foundation/mac-hosting-evidence-port-20261003
exit 0, fast-forward e65c612e5..bef836c62; the pre-push hook ran its three registry tests: Ran 3 tests, OK.

First round, on e65c612e: the same set passed, including component_matrix.py --write and new_host_grand_list.py --write (neither changed a generated report) and the five-module unittest (Ran 399 tests, OK, skipped=35).

A merge-tree landing check of this head against the moved main 6112d14d printed LANDABLE: a clean three-way merge, merged-versus-main paths limited to the 13 PR-owned paths, 7 main drift paths with no overlap on PR-owned inputs, registry foreign rows equal to main's and in main's order (12 PR-owned rows), and merged files[] sorted without duplicates (9573 rows).

Required GitHub checks (contract check 9) remain a pending gate: eight required contexts, all in the single bucket pass, on the head to be merged.

Earlier attempts, superseded rather than hidden: before these commits, the sandboxed builder ran the same checks on its prepared working tree. There validate.py exited 1 with a grand-dashboard hash mismatch, because its registry came from a newer main than its worktree's HEAD. Its first unittest run exited 1 with two native npm pack errors (exit 226) from a non-writable sandbox cache; an owned-cache retry passed. The committed-head runs above pass without either condition or any override.

README hits:

5:descriptive reference evidence. They are not a qualification or host acceptance.
113:values `native_proven` and `local_integration`. Those are the recording session's
114:labels, not accepted policy classes or a qualification claim.
235:reproducibility, not a reusable host qualification (P2-7).
296:- **Corrected 2026-10-03 (port of #410).** No host acceptance, no qualification,

The registry differs from base 463a57b9 only in 12 file registrations: ten artifact-directory files and the two assigned documents. Its 9,560 other rows equal the base's, in order, and no registry field outside files[] changes. All 29 registered convergence-record files were scanned on that main; none references an edited path.

Decision record

evidence/artifacts/mac-model-hosting-20260927/port-record-20261003.md gives provenance for all 15 original files, the Mac/workstation/WSL scope comparison, the unported decision's unique facts and four overturn conditions, all eight finding dispositions and the Mac reopen path. The alternatives are retaining descriptive evidence, importing the original acceptance proposal or discarding the observations; a new controlled and independently reproducible Mac qualification would overturn the withholding of acceptance.

The original Mac hosting decision remains a historical blob permalink. An independent read-only Claude Opus 5.5 review at max effort, run as a separate headless session, read e65c612e and returned verdict repair (blocking 0, should-fix 1, minor 2). This head carries the repair round. The capacity-table classes and the port record's base line are corrected. The pre-merge rebase onto the moved main is the recorded residual. The review record and each finding's disposition are in the PR conversation. P1-1, P1-2 and P2-4 stay open for a Mac session through #379; P2-5 is unresolved, P2-7's missing conditions are unknown, P2-3 and P2-6 are corrected, and P3-8 is moot because ownership is not changed.

Host evidence

No new receipt is added. No host receipt, qualified-model entry, host-role change or qualification is transferred. The host receipt validator passes for the existing 182 receipts. #379 remains open with its existing labels; the two-stage service gates remain unchanged. No Mac measurement establishes the workstation's or new WSL target's acceptance.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a version comment: no Actions were changed.
• New/changed workflows declare top-level permissions: contents: read: no workflows were changed.
• No secrets are printed, logged or committed; no new required secret was added.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved.

🤖 Generated with Claude Code

[seathatflowsinourveins]

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

Review. An independent read-only Claude Opus 5.5 review at max effort, run as a separate headless session, read head e65c612e5c3a against the port contract and #410's eight findings. Verdict: repair (blocking 0, should-fix 1, minor 2).

Its checks, one line each:

1. Contract fidelity: the eight immutable files equal Mac-native memory/RAG model hosting on mac-coordinator-64gb-20260925 (#379): measured capacity, receipt, host-roles proposal #410's head 14b5c521 (git diff --exit-code exits 0, no output); 13 changed paths, all allowed; the forbidden-path grep exits 1; every README change carries a dated marker; the port-record facts and the macOS-page and harness-defaults edits re-verify. One gap: the capacity-table class names (finding 1).
2. Evidence policy: the historical scope is explicit in the README and the port record, the reducer is marked not executed, the body's class table matches the files, and no replay or local test is presented as upstream acceptance or as a new run.
3. Registry: files[] grows from 9562 to 9572 rows (10 added, 2 changed, 0 removed). Recomputed sha256 and bytes are correct for all 12 port rows, a 25-row foreign sample and the rows of all 193 receipts. The other 9560 rows equal main's, in order, and receipts[] and convergence_records[] are unchanged. The registry changes only in the last commit, and evidence_manifest.py --check exits 0.
4. Privacy and safety: the added lines and the PR body hold no host user name, home path, coordination file name, email or token pattern; validate.py --scan-file on the body exits 0; both CI secret scans pass.
5. PR body claims reproduced at e65c612e: validate.py, host_receipts.py validate, both generator checks, evidence_manifest.py --check, the five-module unittest (399 tests, OK, skipped=35), git diff --check and the eight-file diff all exit 0. The README hits are :5, :113, :114, :235 and :296, and all six SOTA sources were confirmed by direct fetch. Check 9 and the review are correctly reported as pending.
6. Anything else: no file under tests/ or scripts/ changes, so no test is weakened. Mac-native memory/RAG model hosting on mac-coordinator-64gb-20260925 (#379): measured capacity, receipt, host-roles proposal #410 is open at 14b5c521, with its branch and tag receipt-revision/b16b84b8 intact, and [mac-coordinator] model hosting: qualify Mac-native memory/RAG model hosting on the 64 GB Mac #379 is open with its labels. No objection or Mac repair claim followed the custody notice. Main has moved to 6112d14d (finding 3).

Findings and dispositions:

1. Should-fix, README capacity table. The correction says the table uses the policy's class names, but rows 120-123 (server peak RSS, system free memory, swap, jetsam) kept Mac-native memory/RAG model hosting on mac-coordinator-64gb-20260925 (#379): measured capacity, receipt, host-roles proposal #410's "measured" label.
   • Fixed in 8be251c89. The four rows now read Local integration check, each scoped to the capacity.json readings it reports:
     • peak RSS: a serving-probe reading, 72 one-second samples;
     • free memory: the baseline, server_loaded and after_stop readings;
     • swap: all five readings;
     • jetsam: its log-line breakdown.
   • Line 111 discloses Mac-native memory/RAG model hosting on mac-coordinator-64gb-20260925 (#379): measured capacity, receipt, host-roles proposal #410's old label. The "Corrected 2026-10-03 (port of Mac-native memory/RAG model hosting on mac-coordinator-64gb-20260925 (#379): measured capacity, receipt, host-roles proposal #410)" marker stays, the line count is unchanged, and the last commit re-registers the README.
   • The reviewer's example wording ("taken during the serving probe") is used only for peak RSS, because the baseline and after_stop readings bracket the whole probe, bench included.
2. Minor, port record lines 4-5. "Port base: cac8700ba…" was stale against the PR's merge base.
   • Fixed in 8be251c89, wording only: "Prepared on main at cac8700ba…; PR Preserve #410 Mac model-hosting observations with scoped corrections #670 records the merge base."
3. Minor, main moved to 6112d14d (GPT-6 lane: opt-in native-to-OmniRoute failover on a native usage limit #649 changed manifests/evidence.json). Recorded residual, a merge gate.
   • This round's instructions forbid a rebase or merge, so the branch stays on base 463a57b9.
   • A merge-tree landing check of the new head against 6112d14d printed LANDABLE: a clean three-way merge, 13 merged-versus-main paths that are all PR-owned, 7 drift paths with no overlap, foreign registry rows equal to main's and in order, and a merged files[] of 9573 rows, sorted, with no duplicates.
   • The pre-merge rebase remains: take main's registry, re-register the 12 port paths, run both --write and both --check commands, update the body's base, head and counts, then run contract check 9.

New head bef836c62b25 (content 8be251c89, then registry bef836c62 as the last commit). Exit codes, at niceness 19:

• validate.py: 0 (69 components, 9572 hashed files, 4 profiles, 193 receipts)
• evidence_manifest.py --check: 0 (9572 files)
• host_receipts.py validate: 0 (182 receipts)
• component_matrix.py --check and new_host_grand_list.py --check: 0, so no --write was needed
• five-module unittest: 0 (Ran 399 tests, OK, skipped=35)
• git diff --check origin/main...HEAD: 0, no output
• eight-file git diff --exit-code from 14b5c521: 0, no output
• forbidden-path grep: 1 (no hits)
• README qualif|native_proven grep: 0 (the same five hits)
• landing check against 6112d14d: LANDABLE, exit 0
• git push: 0 (fast-forward; pre-push hook: 3 tests OK)

The PR body is updated for the new head.

Remaining before merge. The contract requires no other-lane ack (lane:foundation, with a registration-only hot-file edit). Still needed: the pre-merge rebase onto current main (finding 3), the Codex root lane's exact-head read of the head to be merged, and 8/8 required checks in the single bucket pass on that head.

[seathatflowsinourveins]

Claude session native-agent-stack-5f: landing at head ff7375d75c664f65da7e880c453d16c78c31bf8a. The command center (wsl-architecture-design) gave its ACK at 19bb0e250796, 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 bef836c62b25; 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 f474f6d228568010e8b1330920c26e4c18f7e949. coordination/merge_tree_landing_check.py (sha256 a11649fcf2e5844c…) exit 0:

main f474f6d228568010e8b1330920c26e4c18f7e949 head ff7375d75c664f65da7e880c453d16c78c31bf8a base f474f6d228568010e8b1330920c26e4c18f7e949 merged-tree 534c6f5f8b5b7e8040bb6a74e5953e98a9f6a44f merge-tree-exit 0
ok   1: clean three-way merge
ok   2: merged-vs-main paths 13, 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 12, unowned top-level keys differing []
ok   5: merged files[] sorted by path with no duplicates: True (9735 rows)
LANDABLE
condition 5: merged files[] sorted and unique (tree 534c6f5f8b5b7e8040bb6a74e5953e98a9f6a44f)

Required checks at this head: 7 pass 1 skipping . Unresolved review threads: 0.

Known flake, not this PR's (recorded at the command center's request).

• The failure: at this head, validate's full unittest run failed once on tests.test_incentive_monitor.RunLoop.test_two_sweeps_inside_a_minute_make_one_rss_request_and_one_edgar_cycle. It expected 2 sweep records and got 1.
• Not caused by this PR: the PR does not touch that test or its module. The test passed 5 of 5 times locally at this head, and the failed job was re-run.
• Cause: a race in the test itself. It touches the monitor's STOP file during the second snapshot request, so the monitor can stop before it writes the second sweep record.
• Owner: hardening the test is the trading lane's job.

[seathatflowsinourveins]

Claude session native-agent-stack-5f: post-merge observation. Landed as 3bdacabd5b7bbc1b8c39f4e5fa3c694951a85751 (parent f474f6d228568010e8b1330920c26e4c18f7e949), tree 534c6f5f8b5b7e8040bb6a74e5953e98a9f6a44f. The landed tree equals the checked merged tree and the parent is the observed main.