fix(sdk): isolate cache write failures and settle cache telemetry contracts

[rickylabs]

Summary

Implements the approved SDK cache failure-isolation and telemetry-contract plan across three reviewed slices. S1 makes malformed cache evidence fail-safe and bounds namespace cardinality; S2 isolates a real Deno KV oversized cache-write failure after successful data resolution; S3 adds the module-identity diagnostic, mechanically synchronizes its published Query Bridge quotation, and repairs CacheStore mandatory-evidence JSDoc. All three slices passed fresh Tier-A review and the separate-session IMPL-EVAL returned PASS at evaluator commit 0fed4d7ff; the PR is entering exact-head readiness CI.

Scope

• Archetype / area: Archetype 3 runtime-behavior proof profile over packages/sdk (doctrine inventory remains Archetype 2 / Keep)
• Closes fix(sdk): cache write limits must not fail successful query results #1637
• Closes fix(sdk): cache telemetry evidence validation throws into the data path #1619
• Closes fix(sdk): cache namespace cardinality is unbounded because operationId is caller-supplied free text #1620
• Closes SDK cache-provider throw should name its module identity when two SDK instances are loaded #1598
• Closes docs(sdk): CacheStore.get JSDoc documents a return shape that no longer type-checks #1623

Contracts selected and implemented

• fix(sdk): cache write limits must not fail successful query results #1637: isolate only the post-loader store.set() seam. A real Deno KV value-limit rejection records a cache.write provider-error event, returns the already-resolved service payload, and leaves the entry uncached. Loader, lookup, and explicit setCachedData() failures remain fail-loud.
• fix(sdk): cache telemetry evidence validation throws into the data path #1619: malformed provider evidence is fail-safe because observability degrades while application data remains available. Lookup/write return explicitly; invalidation stages each report and merges only after full validation. Descriptor validation runs inside the span after the pending-overflow prologue.
• fix(sdk): cache namespace cardinality is unbounded because operationId is caller-supplied free text #1620: public operationId?: string remains unchanged. A process budget admits 256 normalized namespaces, collapses later values to fixed overflow, and emits the first offending normalized ID exactly once as a span event. Composite construction only syntax-normalizes; admission waits for its first real operation.
• SDK cache-provider throw should name its module identity when two SDK instances are loaded #1598: the diagnostic includes evaluated import.meta.url and presents duplicate SDK module instances as a hypothesis while retaining the browser/client-side hint. Provider ownership remains module-local. The Query Bridge quotation uses <resolved import.meta.url>; the provider test normalizes only that segment and byte-compares the complete stable message.
• docs(sdk): CacheStore.get JSDoc documents a return shape that no longer type-checks #1623: CacheStore.get, set, and delete JSDoc now documents mandatory read/write/invalidation topology evidence. The executed packages/sdk/src/ports/** sweep found no other evidence-contract drift.

Approved D2 contract inversion: the test formerly named “cache telemetry rejects incomplete or unbounded provider evidence” used assertRejects. It now proves that the loader runs and its data returns while the active span records outcome=error and topology_complete=false. This is the approved fail-safe observability contract, documented in packages/sdk/README.md, not a silently weakened guard.

The per-action no-store idea, branded operation IDs, and duplicate-module ownership changes remain deferred.

Slices

• S0 research + Design checkpoint + concrete plan — 89be8da76; terminal PLAN-EVAL artifact cd5193b66
• S1 fail-safe telemetry validation and bounded namespaces — 0e4e26c51, Tier-A PASS
• S2 real Deno KV cache-write failure isolation — 1cf76c6dd, Tier-A PASS
• S3 module diagnostic, synchronized Query Bridge quotation, CacheStore JSDoc, and final gate receipt — 9a26c107a, Tier-A PASS

Validation

• SDK structured check — PASS: 84 files, 1 batch, 0 failed batches, 0 diagnostics
• Focused D4 provider/docs proof — PASS: 1 passed, 0 failed
• Full SDK structured test — PASS: 66 passed, 0 failed
• SDK structured lint / format — PASS: 84 files, 0 findings
• Repo-root structured check — PASS: 2,925 files, 25 batches, 0 failed batches, 0 diagnostics
• Repo-root structured test — PASS: 4,203 passed, 0 failed, 19 ignored; known queue DLQ flake not observed
• deno task quality:scan — PASS: no findings, 7 existing allowances
• deno task arch:check — PASS with existing warnings: exit 0, every package/plugin has FAIL=0
• deno task check:netscript-jsr-specifiers — PASS: scanned=2361 allowances=1 ranges=0 failures=0
• deno task publish:dry-run — PASS: exit 0, including all 12 SDK entrypoints
• SDK JSR audit — PASS with two known warnings: source-directory cardinality and slow-type banner
• Raw deno doc --lint over all 12 SDK entrypoints and separately ./src/cache/mod.ts — both expected exit 1 with the exact six named baseline diagnostics and zero new; neither is represented as green
• deno task surface:diff — RED against the known stale repository baseline: exit 1, 517 undeclared major changes, with the sorted MAJOR set byte-identical at base and head across untouched packages. S3 changes no export map, barrel, TypeScript declaration, dependency, or signature; this red gate is not misrepresented as patch-level.
• Aspire, Docker, and e2e:cli — prohibited and not run

Harness

• Run dir: .llm/runs/fix-sdk-cache-surface-and-telemetry--0.0.7-wave3/
• Phase: implementation complete — S1, S2, and S3 Tier-A passed; separate-session IMPL-EVAL PASS
• PLAN-EVAL: terminal PASS at evaluated head ee1b44c6d, artifact cd5193b66
• Scope: implementation stayed within each coordinator-authorized slice; S3 changed exactly the provider message, CacheStore JSDoc, adjacent provider test, authorized Query Bridge page, and run artifacts
• Merge only after required exact-head readiness CI is terminal and truthful.

Drift / Debt

• The brief uses the stricter Archetype 3 proof profile while the current doctrine inventory classifies SDK as Archetype 2 / Keep.
• Full SDK/cache raw doc lint remains pre-existing red with the accepted exact six-diagnostic, zero-new baseline.
• Repository-wide surface:diff remains unusable as a leaf patch-level verdict because its checked-in baseline reports 517 undeclared major changes with an identical sorted MAJOR set at base and head across untouched packages; this is recorded as a red limitation, not hidden.
• The Query Bridge quotation is synchronized in place and mechanically protected; no documentation debt remains for D4.

Definition of Done

• Topic orchestrator authorized exactly the five named additional paths and accepted the exact six-diagnostic doc-lint baseline.
• Separate PLAN-EVAL returns PASS before any product implementation.
• The real Deno KV oversized-value test is demonstrated RED on the pre-fix behavior and GREEN after isolation.
• All five issue contracts are implemented without widening the public TypeScript surface.
• Repo-root structured check/test plus SDK lint/fmt, JSR, publish, quality, architecture, exact-pin, and surface gates satisfy the approved bar.
• Separate-session IMPL-EVAL returns PASS.

[rickylabs]

[PHASE: RESEARCH]

Re-baselined the five issues against main@baf1cdf67.

Findings

1. fetchAndCache() awaits store.set() after the service payload resolves and rethrows persistence failure, so the current data-path defect is confirmed.
2. Lookup, write, and invalidation can all reach the shared typed-never malformed-evidence throw.
3. Descriptor validation is evaluated as a withSpan argument, so an invalid descriptor currently throws before any span exists.
4. Static generated resource.action and composite namespaces are pinned by tests; public direct operationId remains free text.
5. Whole-repo census found SDK internal consumers, Fresh public consumers, the KV backend edge, CLI generated-import assertions, MCP generated docs, and release/docs tooling. The gate plan therefore uses repo-root test.
6. Ports JSDoc sweep found only CacheStore get/set/delete drift.
7. SDK publish dry run and exact-specifier guard pass; full-export doc lint is pre-existing red and is recorded as a scope ruling.

Artifacts

• research.md
• scope-boundary.md
• drift.md

Next

Topic orchestrator reviews the scope boundary and sends the plan to a separate PLAN-EVAL session. No implementation has started.

[rickylabs]

[PHASE: PLAN]

Tier-A plan repairs are committed through ee1b44c6d; implementation remains blocked pending coordinator-owned PLAN-EVAL.

Final repairs

1. T-1 — overflow deferral: CacheQuery and provider-boundary calls carry an invocation-local admission decision into a first-statement span prologue. The prologue emits the pending overflow event and validates the descriptor inside the span. composite-query.ts:42 syntax-normalizes only; admission and emission wait for its first real spanned operation.
2. T-2 — invalidation rollback: each report is staged and merged only after complete validation. A rejected report emits none of its earlier partial entries; valid reports may remain, while the final span stays incomplete. The plan requires a third-entry failure test.
3. T-3 — named doc baseline: raw combined-12-entrypoint and cache-entrypoint deno doc --lint commands both remain expected exit 1 with exactly six named diagnostics and zero new. Neither is a PASS.
4. T-4 — synchronized published quote: the coordinator superseded the prior scope treatment. docs/site/web-layer/query-bridge.md is the fifth authorized path. Its single-line quotation uses <resolved import.meta.url> for the install-specific segment; cache-provider_test.ts normalizes only that segment and byte-compares the full stable message—including [NetScript SDK] —against runtime.

Exact additional surface

1. packages/sdk/README.md
2. packages/sdk/src/cache/cache-provider_test.ts
3. packages/sdk/tests/cache/cache-telemetry_test.ts
4. packages/sdk/tests/cache/cache-query-kv-limit_test.ts (new)
5. docs/site/web-layer/query-bridge.md

No other path is authorized. No other site documentation will be swept or edited.

Evidence

• Latest commit: ee1b44c6d401a9edb9c8690870ea2d9151f8f504
• Only six run artifacts changed in this plan repair; no product/site implementation was performed.
• git diff --check passed.
• Draft state and sole status:plan lifecycle state are retained.

Next

Coordinator-owned PLAN-EVAL only. Do not begin implementation, dispatch an evaluator, or move the PR out of draft.

[rickylabs]

PLAN-EVAL — PASS

Evaluator: native Claude claude-fable-5 · effort medium · Remote Control (session_01GaNTjv6oY6MaxnKHH1ZfrB) · job 0287ccbe · route matched (respawnFlags) · separate session from Codex author 01a00516….
Evaluated head: ee1b44c6d401a9edb9c8690870ea2d9151f8f504 · base baf1cdf67.
Artifact: .llm/runs/fix-sdk-cache-surface-and-telemetry--0.0.7-wave3/plan-eval.md @ cd5193b66.

Item	Result	Evidence
Phase invariant	holds	git diff --name-only base..HEAD → 8 paths, all .llm/runs/
T-1 overflow deferral	confirmed repaired	11 calls + def censused (cache-query.ts:85,305,329,361,389, cache-provider.ts:125,141,159,169,176, composite-query.ts:42); every executable site has a withSpan callback; composite factory intentionally not admitted; single prologue (flush → validate) makes ordering coherent
T-2 non-throwing recorder	confirmed hole; plan closes it	recordCacheWrite falls through into loops; recordCacheInvalidation banks entries pre-throw and ends with unconditional topologyComplete:true (cache-telemetry.ts:339-401); per-report staging + third-entry-failure test is a real proof
T-3 doc-lint baseline	confirmed	ran 9a/9b at head: exit 1 / exit 1, exactly the six pinned private-type-ref diagnostics by symbol+location; nowhere described as pass; named comparison, no count verdict
T-4 Query Bridge fifth path	confirmed	FOLLOWUP-DOC-QUERY-BRIDGE-DIAGNOSTIC removed in ee1b44c6d, no residue; single-token normalization is unambiguous; test-only docs read
Scope	exact 4+5, no sixth	query-options.ts untouched; every D1–D5 mechanism fits
D1–D5	feasible	getKv({provider,path,skipServiceDiscovery}) keys exist and singleton is reused by KvCacheStore; RED drives real Deno.openKv(':memory:') value-size rejection; D2 amend preserves the span guard; D3 bounded 256+bool; D4 ownership unchanged; D5 JSDoc-only
Validation honesty	ok	root test/check cover asserting packages (Fresh, CLI, KV, MCP, surface baseline); no Aspire/Docker/e2e

Non-blocking advisories: (1) try/finally around the D3 registry reset; (2) resetKv()/closeKv() teardown in the KV-limit test; (3) keep the single-line docs block unwrapped; (4) keep internal test-reset/admission helpers off cache/mod.ts.

Implementation may begin per the slice table; IMPL-EVAL remains a separate mandatory session.

[rickylabs]

IMPL — S1 progress

Implemented the coordinator-authorized S1 only at 0e4e26c51: fail-safe cache evidence validation (D2) and bounded namespace admission/span prologue (D3).

Scope

Changed only the authorized S1 product/test/doc files plus run artifacts:

• packages/sdk/src/cache/cache-telemetry.ts
• packages/sdk/src/cache/cache-query.ts
• packages/sdk/src/cache/cache-provider.ts
• packages/sdk/tests/cache/cache-telemetry_test.ts
• packages/sdk/README.md
• .llm/runs/fix-sdk-cache-surface-and-telemetry--0.0.7-wave3/{context-pack.md,worklog.md}

S2 and S3 are untouched. In particular, there is no KV-limit test, no CacheStore/provider-diagnostic change, and no Query Bridge edit.

Behavioral proofs

All six required proofs live in packages/sdk/tests/cache/cache-telemetry_test.ts:

1. Partway invalidation rollback — the rejected report banks valid L1/L2 entries, fails on its third invalidation, emits neither staged tier, preserves a separate valid durable report, and ends incomplete/error.
2. Normal data return with malformed evidence — malformed lookup and write reports preserve the loader payload; malformed invalidation completes normally while recording incomplete/error.
3. Request-local first overflow — after 256 admitted ids, the 257th collapses to fixed overflow, names the normalized offender exactly once in a span event, later ids collapse without raw-id retention/emission, and console.warn is never called.
4. Composite construction then operation — construction creates no span and consumes no budget; its first real operation performs admission and emits the pending overflow event.
5. try/finally reset — both budget-filling tests reset the process-global namespace registry in finally; the composite test also resets the cache provider.
6. Descriptor validation in-span — an invalid descriptor still returns loader data; the recorded event order proves overflow flush happens before descriptor validation.

Invalidation aggregation now stages each report in a private map and merges it only after the complete report validates. Internal admission/reset/prologue helpers remain off packages/sdk/src/cache/mod.ts and packages/sdk/mod.ts; their barrel diff is empty.

Validation

• SDK structured check: PASS — 83 files, 0 failed batches, 0 diagnostics.
• Telemetry focused test wrapper: PASS — 21 passed, 0 failed.
• Expanded focused cache/query wrapper: PASS — 32 passed, 0 failed.
• SDK structured lint: PASS — 83 files, 0 findings.
• SDK structured format: PASS — 83 files, 0 findings.
• deno task quality:scan: PASS — no findings; 7 existing allowances.
• deno task arch:check: PASS with existing warnings — 0 failures. Both changed cache sources are below the 500-line cap; the existing SDK source-directory cardinality warning remains.
• Supplemental repository-wide surface:diff: FAIL / not an S1 verdict — exit 1 against a stale workspace baseline, reporting widespread signature drift in untouched packages. This is not represented as a pass. The S1-specific export proof is the empty cache/root barrel diff.

Raw deno doc --lint was not rerun in this focused slice. No root test, Aspire, Docker, or CLI E2E gate was run.

Carried advisories

• S2 must use resetKv() / closeKv() teardown for the process-global @netscript/kv singleton.
• S3's single-line Query Bridge fenced quotation must remain unwrapped.
• The namespace admission/reset/prologue helpers must remain off all published barrels.

The PR remains draft with its sole status:plan label. No evaluator was launched or arranged; implementation stops here pending the coordinator's fresh Tier-A review.

[rickylabs]

[PHASE: IMPL]

S2 is committed and pushed at 1cf76c6dd691378eddbbd9cd3c8a82d50c30fa2f. This slice implements D1 only.

Scope

Changed exactly:

• packages/sdk/src/cache/cache-query.ts
• packages/sdk/tests/cache/cache-query-kv-limit_test.ts (new)
• S2 run artifacts under .llm/runs/fix-sdk-cache-surface-and-telemetry--0.0.7-wave3/

No S1 or S3 file changed.

D1 behavior

Only the post-loader store.set() rejection seam is isolated. It records recordCacheProviderError(span, CacheOperations.READ, namespace, descriptor, CacheEvents.WRITE) and returns the already-resolved data. Cache lookup failures, loader failures, background refresh behavior, and explicit setCachedData() remain fail-loud.

The proof uses a real :memory: Deno KV and real KvCacheStore, not a synthetic throwing store. It asserts:

1. the caller receives the identical resolved payload;
2. the loader runs once;
3. the read span contains a cache.write event with outcome=error and topology_complete=false;
4. a follow-up cache lookup misses;
5. explicit setCachedData() still throws the real KV size error;
6. finally awaits both closeKv() and resetKv().

RED / GREEN

RED before the product edit:

{"schemaVersion":1,"command":["deno","test","--reporter=tap","--allow-all","packages/sdk/tests/cache/cache-query-kv-limit_test.ts"],"cwd":"/home/codex/repos/netscript-007-leaf-sdk-cache","exitCode":1,"durationMs":4511,"summary":{"passed":0,"failed":1,"ignored":0,"totalResults":1,"uniqueFailures":1},"failures":[{"message":"TypeError: Value too large (max 65536 bytes) ... at async KvCacheStore.set ... at async CacheQuery.fetchAndCache","count":1,"tests":[{"name":"CacheQuery returns loaded data when real Deno KV rejects an oversized cache write","file":"./packages/sdk/tests/cache/cache-query-kv-limit_test.ts","line":50}]}]}

GREEN on the pushed implementation:

{"schemaVersion":1,"command":["deno","test","--reporter=tap","--allow-all","packages/sdk/tests/cache/cache-query-kv-limit_test.ts"],"cwd":"/home/codex/repos/netscript-007-leaf-sdk-cache","exitCode":0,"durationMs":1663,"summary":{"passed":1,"failed":0,"ignored":0,"totalResults":1,"uniqueFailures":0},"failures":[]}

The unabridged structured RED and GREEN are committed in s2-report.md.

Final gate verdicts

• Focused SDK check: PASS — 84 files, 1 batch, 0 failed batches, 0 diagnostics.
• Real-KV focused test: PASS — 1 passed, 0 failed.
• Full SDK test: PASS — 66 passed, 0 failed.
• SDK lint: PASS — 84 files, 0 findings.
• SDK format: first run FAIL with one new-test import-format finding; corrected and final run PASS — 84 files, 0 findings.
• Repo-root deno task check: PASS — 2,925 files, 25 batches, 0 failed batches, 0 diagnostics.
• Repo-root deno task test: PASS — 4,203 passed, 0 failed, 19 ignored, 4,222 total.
• deno task quality:gate: PASS with existing warnings — quality scan has no findings; architecture check has zero failures.
• Raw deno doc --lint: not run in S2.
• Aspire, Docker, and e2e:cli: not run, as prohibited.

The PR body now explicitly explains S1's approved D2 assertRejects contract inversion. No checkbox was ticked, and the PR remains draft with its sole status:plan label.

S2 stops here pending the coordinator's fresh Tier-A review. No evaluator or review was launched or arranged.

[rickylabs]

[PHASE: IMPL] [SLICE: S3]

S3 is implemented and pushed at 9a26c107afa75bf1f38b78fe96c6df533b156c36. The PR remains draft and awaits fresh Tier-A review.

Scope

Changed exactly the S3-authorized product/test/docs paths plus run artifacts:

• packages/sdk/src/cache/cache-provider.ts — message wording only
• packages/sdk/src/ports/cache-store.ts — JSDoc only
• packages/sdk/src/cache/cache-provider_test.ts
• docs/site/web-layer/query-bridge.md
• .llm/runs/fix-sdk-cache-surface-and-telemetry--0.0.7-wave3/{context-pack.md,worklog.md,s3-report.md}

S1/S2 files remain byte-identical. _provider remains module-local; runtime code adds no file/asset read.

Behavioral/documentation proof

• D4 focused wrapper: 1 passed, 0 failed. The test captures the real error, proves its URL equals new URL('./cache-provider.ts', import.meta.url).href, normalizes only that URL to <resolved import.meta.url>, and byte-compares the complete single-line Query Bridge fence.
• D5 whole-ports executed sweep found no evidence-contract drift outside cache-store.ts; the three repaired comments document read, ordered write/promotion, and invalidation topology evidence.

Gates

• SDK check: PASS, 84 files / 0 diagnostics
• SDK tests: PASS, 66/0
• SDK lint/fmt: PASS, 84 files / 0 findings
• Root check: PASS, 2,925 files / 0 diagnostics
• Root test: PASS, 4,203 passed / 0 failed / 19 ignored; known queue flake not observed
• Quality scan: PASS, no findings
• Architecture check: PASS with existing warnings, exit 0 and all FAIL=0
• Exact NetScript specifiers: PASS, scanned=2361 allowances=1 ranges=0 failures=0
• Publish dry run: PASS, exit 0 including all SDK entrypoints
• SDK JSR audit: PASS with the two known warnings (source cardinality and slow-type banner)
• Raw doc lint: expected RED, both exit 1 with exactly the six pinned named diagnostics and zero new; neither is called a pass
• Surface diff: RED, exit 1 against the known stale repo baseline (524 undeclared major changes across untouched packages). S3 itself changes no export map, barrel, declaration, dependency, or signature.

Full evidence: .llm/runs/fix-sdk-cache-surface-and-telemetry--0.0.7-wave3/s3-report.md.

Next

Coordinator-owned fresh Tier-A review. This implementation session did not launch or arrange an evaluator, did not run Aspire/Docker/e2e:cli, and did not change readiness, labels, milestone, or issue state.

[rickylabs]

Delta IMPL-EVAL — generated-asset repair 7549d9fc0 — PASS

Evaluator: fresh native Claude session (claude-fable-5 · medium · Remote Control) — job 08eb7184, session_01Jc8aRcLQFVyVWKogq6SaFC; separate from the Codex author and the Tier-A orchestrator. Route matched. Evaluated head 7549d9fc0 (== PR head at eval time).

Scope: source→generated fidelity + confinement of 0fed4d7ff..7549d9fc0 only. Prior product IMPL-EVAL PASS at 9a26c107a preserved (git diff 9a26c107a..7549d9fc0 -- packages/ plugins/ docs/ tools/ deno.json deno.lock is empty).

Item	Result
1. Confinement	✅ only prose.json.gz, provenance.json, run artifacts
2. Fidelity	✅ deno task check:agent-docs-prose exit 0, fresh:true, stalePaths:[]; --check byte-compares fresh corpus vs asset; mutating regen at HEAD left tree clean (byte-identical). files manifest unchanged (181). Committed payload 4753909 B / sha 6df99eb8…853fe; gz 1363396 B. Entry-level diff vs old corpus: only pages/web-layer/query-bridge/index.md + llms-full.txt changed
3. Diagnostic text	✅ <resolved import.meta.url> 0→4 occurrences; single-line template present in the query-bridge entry
4. Product preservation	✅ empty diff
5. Baseline reds	✅ still red, unchanged: F-DOCT-5 13 children; surface:diff exit 1 (517 undeclared); doc-lint 9a exit 1 (3 diagnostics), 9b exit 1 (3 diagnostics), exact names/locations match

Supervisor claims: all CONFIRMED (incl. lint 0 / fmt 0 over 84 files).

Blocking: none. Advisories (non-blocking): provenance.sourceCommit = rendered tree commit 0fed4d7ff, not the asset commit — by generator design.

Artifact: .llm/runs/fix-sdk-cache-surface-and-telemetry--0.0.7-wave3/delta-eval-generated-assets.md @ 72d57229f.

[rickylabs]

[PHASE: IMPL]

Second post-eval generated-asset repair pushed at 27a64ea4c122312e46ffc9602509c191bcd0dd71.

This closes the third tracked link in the chain:

docs/site/web-layer/query-bridge.md
  -> .llm/assets/agent-docs/**
  -> packages/cli/src/kernel/assets/agent-docs.generated.ts

The barrel was regenerated only through deno task gen:assets-barrel. Immediate confinement proof:

 M packages/cli/src/kernel/assets/agent-docs.generated.ts

The commit contains that one generated file plus three artifacts in the existing harness run directory. The other six barrel outputs, canonical prose assets, site sources, hand-written package/plugin files, and lockfiles did not change.

Validation

• deno task check:assets-barrel: PASS, exit 0; all seven targets reproduce and the generated-file diff is clean.
• deno task check:agent-docs-prose: PASS, exit 0; fresh=true, stalePaths=[].
• CLI structured check: PASS, exit 0; 883 files, 8 batches, 0 diagnostics.
• CLI lint wrapper: NOT A PASS, exit 2; Deno excludes the relevant CLI targets. The wrapper reports 0 lint occurrences but correctly refuses a false green. An explicit touched-file run is also excluded.
• CLI format wrapper: NOT A PASS, exit 2 for the same exclusion; 0 format findings, but no valid formatter verdict. The canonical generator formats through deno fmt --ext ts -, and asset regeneration is byte-stable.
• Combined/cache raw doc-lint: EXPECTED BASELINE RED, exit 1 each, exactly the pinned three-plus-three diagnostics and zero new.

Further generated link found

The downstream audit found a fourth stale generated link outside this repair's authorization:

{"gate":"deno task check:publish-assets","exitCode":1,"stalePaths":["packages/mcp/src/publish-assets.generated.ts"]}

That file is a sibling consumer of .llm/assets/agent-docs/provenance.json: it still embeds sourceCommit: 504de3f67 while the canonical provenance is 0fed4d7ff. It was checked read-only and left untouched. Whole-repo generator/consumer searches found no fifth checked-in generated artifact after the MCP file.

Detailed evidence: .llm/runs/fix-sdk-cache-surface-and-telemetry--0.0.7-wave3/assets-barrel-repair-report.md.

No root suite, Aspire, Docker, CLI E2E, evaluator, or review dispatch was run. Stopping for the coordinator's scope ruling, fresh Tier-A, and asset-chain delta evaluation.

[rickylabs]

Generated-asset repair receipt — link 4/4

Pushed 9a2c74c41990c1e2a56c9714834fff97feb63466. This closes the fourth and final checked-in link of the authorized cascade:

docs/site/web-layer/query-bridge.md
  → .llm/assets/agent-docs/**
    → packages/cli/src/kernel/assets/agent-docs.generated.ts
      → packages/mcp/src/publish-assets.generated.ts

The canonical deno task gen:publish-assets task changed exactly one product path, packages/mcp/src/publish-assets.generated.ts, plus run artifacts. Its one-line generated delta advances MCP_EMBEDDED_DOCS_PROVENANCE.sourceCommit from 504de3f67 to 0fed4d7ff; the selected twelve-document payload and digest are unchanged.

Final-head cascade gates:

• check:publish-assets: PASS (exit 0)
• check:assets-barrel: PASS (exit 0)
• check:agent-docs-prose: PASS (exit 0, fresh: true, stalePaths: [])
• MCP check wrapper: PASS (115 files, 0 diagnostics)

The unfiltered MCP lint/fmt wrappers are honestly red at config discovery (exit 1, zero findings) because the root selection enters deliberately invalid doctor fixture workspaces. The same structured wrappers excluding only packages/mcp/tests/fixtures/doctor/** pass 110 files with zero lint or format findings. Both raw SDK doc-lint commands remain exit 1 with exactly the pinned three-plus-three baseline diagnostics and zero new diagnostics.

An executed generator/consumer search found no generator downstream of the MCP generated file; no fifth checked-in mirror was found. No evaluator or review was launched or arranged.

[rickylabs]

Delta IMPL-EVAL — generated-asset chain: PASS

Field	Value
Evaluated head	9a2c74c41990c1e2a56c9714834fff97feb63466 (merge base baf1cdf67)
Evaluator	fresh native Claude session claude-fable-5 · effort medium · Remote Control (session_01E3QfD1wkvb1naZKS6m7bp2), separate from Codex author 01a00516-… and the Tier-A orchestrator
Artifact	.llm/runs/fix-sdk-cache-surface-and-telemetry--0.0.7-wave3/delta-eval-asset-chain.md @ ac274a464
Scope	four-link chain only (query-bridge.md → .llm/assets/agent-docs/* → agent-docs.generated.ts → publish-assets.generated.ts); product PASS 9a26c107a and corpus PASS 7549d9fc0 preserved

Findings (all executed independently in a detached worktree at 9a2c74c41):

1. Simultaneous freshness — check:agent-docs-prose EXIT 0 (fresh:true, stalePaths:[], sha 6df99eb8…), check:assets-barrel EXIT 0, check:publish-assets EXIT 0, git status --porcelain empty afterwards. Confirmed.
2. Convergence — barrel payload decoded independently: 4 753 909 bytes, sha256 6df99eb856eb…853fe = provenance.json = live check output; sourceCommit 0fed4d7ff identical across links 2/3/4; link 3 diff 6+/6−, link 4 diff 1+/1− (sourceCommit). Confirmed.
3. Content fidelity — Cache provider not initialized in module <resolved import.meta.url> present ×2 in prose.json.gz and in the decoded CLI barrel (0 at base); absent from the MCP file because MCP_EMBEDDED_DOC_PATHS excludes query-bridge. Confirmed.
4. No fifth mirror — every gen:*/check:* run at both heads: check:mcp-export-corpus EXIT 1 at both heads with an identical 5+/5− regeneration diff (pre-existing); emitted-samples, streams-types, scaffold-versions, netscript-jsr-specifiers, aspire-host-ports EXIT 0 at both. No new link.
5. Source movement — baf1cdf67..HEAD touches exactly the eight S1–S3 files + query-bridge.md + four chain links. Confirmed.
6. Pre-existing reds still red — surface:diff EXIT 1 (965 lines, same at base), F-DOCT-5 13 children in packages/sdk/src, check:mcp-export-corpus EXIT 1, six deno doc --lint private-type-ref diagnostics. None silently fixed.

Coverage limits confirmed: run-deno-lint.ts --root packages/mcp errors without a verdict (direct deno lint on the generated file: clean); agent-docs.generated.ts is lint/fmt-excluded, so it rests on check:assets-barrel byte-equality (green). #1652's derivedAssetCascadePaths = the same four paths (milestone-cluster-state.json:512-517, PR files) — precedent confirmed.

Blocking items: none. Advisories (non-blocking): stale mcp-export-corpus on main (separate chore); run-deno-lint.ts per-package workspace-config gap.