Restore bounded cold broad retrieval on large installed publications

[TheGreenCedar]

CodeStory's installed-host broad search can fail immediately after a complete Bazel publication. This prevents the original adaptive P04-Q12 investigation (“How does remote execution exchange action inputs and outputs?”) from reaching an answer-model interaction.

At unpublished dev candidate 5b38104c with pinned Bazel source 1f05f210, two separate fresh installed-host phase-one attempts reached a full core and sidecar publication, then the same first search returned retrieval_unavailable from retrieval_engine with a lexical stage_deadline and zero candidates. In the second attempt, one exact-query retry failed identically (1,159–1,160 ms lexical trace). The host RPC completed within 20 seconds; resource and native supervision stayed within the 12-GiB phase cap and 50-GiB reserve, with authenticated drain. No answering-model session started. Preserve both failed attempt receipts and closed budgets. The direct-CLI full-retrieval success belongs to a separate publication and does not prove the installed-host interaction.

Investigate the owning retrieval/selection boundary and reproduce the cold large-repository failure before changing source. Repair the smallest general behavior that makes the existing public search usable without weakening fail-closed publication or cancellation semantics, inflating the caller's fixed deadline, or adding Bazel/question-specific selection. Verify the causal failure and credible cancellation/concurrency boundaries with focused tests. Independent review must see the exact patch. Then rebuild the affected unpublished candidate and return to the original installed-host P04-Q12 interaction under a fresh supervised attempt.

Refs #2275, #2173, #2135. This child issue does not close those parents or establish answer sufficiency, comparative gain, acceptance, or release readiness.

[TheGreenCedar]

Active source lane: codex/2370-bazel-cold-retrieval, isolated worktree ~/.codex/worktrees/0176-bazel-cold-retrieval/CodeStory, base and current head 5b38104c5d7e7336d7e98d86dc9d678be28ae988. A Sol 6 implementer owns the causal retrieval patch and focused tests; a separate Sol 6 challenger owns the hostile contract and exact-patch review. Root owns Git integration, package installation, native execution and the return to original P04-Q12. Cargo is reserved to the implementer serially.

The blocking observation is two closed fresh installed-host phase-one failures, with a same-query retry in the second. This issue's next proof is a red causal reproduction, a bounded retrieval fix with cancellation/publication hostile checks, and independent patch acceptance. No new candidate build or model run has started for this source slice. Parent #2275 and unified coordinator/breadth remain open.

[TheGreenCedar]

Source repair is draft PR #2371 at 4a3d87623c7ef4ecf8329c88cbdeee77479adba7 (base dev 5b38104c5d7e7336d7e98d86dc9d678be28ae988). The regression was red on the base and green on the patch. Focused lexical, executor, cancellation, formatting and diff checks passed. Independent exact-commit source review accepted the one-file patch with no concrete counterexample; its retained receipt is 2370-bazel-cold-retrieval-review/source-review.json, SHA-256 1a06d0cdfb8d46b9c81058ccbc7a31e287cd05da1ec2e7dd22115cecedcd3613.

CI is still pending. This is source-level acceptance only. The two installed-host Bazel stage-deadline failures and their closed budgets remain unchanged. Next proof, after eligible integration, is a new candidate and the original installed-host P04-Q12 investigation under fresh supervised controls. Keep #2275, #2173 and #2135 open; answer sufficiency and release readiness are not established.

[TheGreenCedar]

PR #2371 merged into dev as c3a5a7378ecdf57440312e433aaf2df8a8e21dc9, tree 639695ff95e06de36c1b4ab871cfdd0d118fb495, after exact-commit Sol 6 source review and all applicable draft checks passed. The one-file fix defers lexical body/source reads until admission without changing the 4,096 candidate boundary. This is accepted source, not installed-host acceptance.

The new unpublished candidate is being prepared from a clean detached checkout at that exact dev head. The original Bazel P04-Q12 installed-host search and full adaptive answer remain open. The -01 and -02 stage-deadline failures retain their closed budgets. Keep this child issue open until its motivating installed-host interaction is checked; #2275, #2173, and #2135 remain open.

[TheGreenCedar]

The c3a5 source fix is merged and its unpublished package plus isolated installation have independent package/control acceptance. I reopened this child issue because the PR merge auto-closed it before the required installed-host interaction was accepted.

Fresh installed Bazel attempt bazel-q12-adaptive-devc3a5-01 is closed as a control failure, with its 900-second/12-GiB budget closed. The first public search on pinned Bazel 1f05f210 returned from a complete publication with the authenticated c3a5 runtime and ten evidence items. The following observational status returned ready, live_ready=true, and retrieval_mode=full from the same runtime, but correctly used served_from=contract_only and carried no publication generation. The test control incorrectly required a complete publication generation on status, so no phase-one acceptance or answer-model interaction was recorded. The source-level search improvement is a useful diagnostic observation, not yet accepted installed-host or answer-sufficiency proof.

The failed supervisor and raw transcript are preserved at bazel-q12-adaptive-devc3a5-01/root-failure-decision.json (SHA-256 aec3bc58257633cbdf498dc033afd13f356e9ef1efeee4954b48fe618ddb9a04). Native drain was authenticated; sampled peak added disk was 3,508,371,456 bytes, under the 12-GiB cap. Only owned mutable attempt files were removed after handle/ownership checks; the cleanup receipt is root-owned-cleanup.json (SHA-256 5a602d70f4d70bbd76a4d33b218179a2884443198b740ad9b60ce0b13fe33e91).

A distinct fresh attempt is being prepared with independent review of separate search-publication and observational-status predicates. No model run or release claim follows from this failure. Parent #2275, breadth #2173 and coordinator #2135 remain open.

[TheGreenCedar]

Fresh installed-host diagnostic bazel-q12-adaptive-devc3a5-02 completed under independently accepted controls. The exact c3a5 runtime and pinned Bazel source closed cleanly; authenticated native drain and scoped cleanup completed in 232.74 seconds with 3,508,756,480 bytes sampled peak added disk, below the 12-GiB stop cap. No model session ran. Root decision: bazel-q12-adaptive-devc3a5-02/root-phase1-decision.json, SHA-256 ebb748d4abf07ae270ca8aeb62124ada342e641a088ec5d08a94ad9afa486c5b. This fresh attempt budget is closed.

The first cold public search still failed with typed retrieval_engine / lexical stage_deadline. Its single allowed same-query retry succeeded from a complete publication with the authenticated c3a5 runtime; status separately reported live/full readiness. This establishes bounded retry recovery for this installed path, not first-call reliability or answer sufficiency. The ten returned evidence items were near-duplicate versioned CLI reference pages, so the original remote-execution investigation still needs an adaptive interaction and quality judgment.

The previous -01 control failure remains closed and separate. Next: a fresh original Bazel Q12 adaptive investigation under its own reviewed admission, then causal product decision. Keep this child and #2275/#2173/#2135 open.

[TheGreenCedar]

Fresh c3a5 installed-host Bazel Q12 attempt bazel-q12-adaptive-devc3a5-03 reached a completed Luna 5.6-medium participant turn, but the supervised attempt failed after model completion. A stale already-acknowledged boundary request was checked against the pre-model cutoff before the duplicate-ID return. The attempt budget is closed; raw rollout, answer, readiness transcript, native closeout and scoped cleanup are preserved. Root failure receipt SHA-256: a6ac05564dddd2dc56351264aabffc9c4020002351168ac90b380d9e899b5621.

Independent development grading of that preserved answer is 46/70, below the frozen 49/70 threshold, with no critical unsupported claims. This is development evidence inside a failed supervised run, not a panel pass. CodeStory's contribution was limited: the first available packet emphasized repeated versioned command-line docs; six later public searches had full/available retrieval and some relevant source paths, but also null excerpts and unrelated or repeated results. The participant relied heavily on ordinary source reads and omitted worker-side details that were available there. Grade receipt SHA-256: 6abef2fdf705729383a361bf71667d380ba1d83fb945900bfd7dca4a679e7f2d; append-only correction SHA-256: f910807ee49da5da22f94c1ec1da5241b53bf188f7679ce7e2a2c5f293ddcb9d.

Root product decision: preserve the merged lexical deferral and accepted package, keep this child and #2275/#2173/#2135 open, and isolate the evidence-selection/presentation gap from host-agent omission using the recorded results before another costly full interaction. First-call reliability and adaptive answer sufficiency remain unproved. Fix the supervisor's duplicate-boundary ordering before reusing that control, without reopening any closed budget. Root decision receipt SHA-256: 89a171a379c1387d68eee3103eb314a9d5f3177c392fce2a95363a37abde53cb.

[TheGreenCedar]

Correction to my preceding Bazel Q12 status comment: the retained participant commands show client-side executor, cache and output source reads, but no worker ExecutionServer or equivalent worker-implementation read. CodeStory did not usefully surface the worker flow in the packet or six follow-up searches. The agent's answer omitted some client-side cache and output qualifiers it had read, while worker-side Action/Command fetch, input materialization, command execution and output upload remained an unresolved navigation/evidence gap. The independent evidence-allocation receipt is development-grade-q12-evidence-allocation.json, SHA-256 bd64e2c0ff963a9ca4b1f0ee87540d98150b7ae08adddde7c457206febc25560; root's append-only decision correction SHA-256 is 68baf984fa356854da0b9099910b641bf30a0f29304608c326ac04f001b43074.

This correction does not change the 46/70 development grade, lack of critical unsupported claims, limited CodeStory contribution, failed-supervision classification, or closed budget. The next product check must discriminate whether existing public operations can lead from the client remote-execution path to relevant worker implementation with usable excerpts; it must not assume the agent already saw those worker facts.

[TheGreenCedar]

Q12 worker-side reference is now pinned without reopening the failed panel. At Bazel 1f05f210, ExecutionServer.execute downloads the Action/Command and input tree, runs in the materialized execution root, uploads an UploadManifest, and returns an ActionResult through ExecuteResponse in the completed Operation. Exact source windows: src/tools/remote/src/main/java/com/google/devtools/build/remote/worker/ExecutionServer.java:184-200,287-378,409-455.

This completes the engineering reference path, not the participant answer. The attempt03 participant did not read the worker source, and its CodeStory calls did not supply that evidence. The smallest product follow-up is one fresh project-scoped context query for ExecutionServer, then snippet/source inspection of those windows under a new admitted budget. The current host does not expose CodeStory MCP tools, so no installed-host follow-up was run here. No source fix is inferred from a source-only reference. Local receipt: bazel-q12-adaptive-devc3a5-03/worker-side-source-reference.json, SHA-256 1c03610bdba6d6fb609eb984f9cae0319e8cb124cf032dc738319af94617ec27.