Validate live bounded S3 and W&B model-run collection #5

Windows primary-checkout delivery — 2026-09-27

The response-guard candidate from the archived package was verified (34/34 manifest entries) and integrated on existing draft PR #1 as verified SSH-signed commit 5a7e4418bcebee9c9ceccca4837e04f881d57ecf. The change guards W&B hasNextPage/continuation metadata before accepting a page and closes an opened response if header parsing fails. It preserves valid earlier pages and existing request/byte budgets.

On Python 3.13.12, the focused collector/cache/CI-prerequisite/pipeline suite reported 110 passed, 1 skipped (fresh real export unavailable). Critical Ruff E9,F, Python compilation, and Git whitespace checks passed. These are source tests with synthetic pages, not authorized live W&B/S3, a fresh producer-to-UI run, full locked backend acceptance, or Jev parity. The full Ruff rule set reports pre-existing findings in this branch; the prior hosted backend job stopped before lint/tests for missing authorized private-index token. Keep this issue's live acceptance boxes unchecked and PR #1 draft.

Confirmed source defects and uncommitted repair candidate — 2026-09-27

The single existing integration PR #1 remains draft at ed825bf1c2c4c19eb3de99ad46564c5803bb4a39. Independent source review matched the current provider to Git blob 19c5d510130537a6153bcafdc1ec4cc56b6b6dab and reproduced two defects:

1. A malformed pageInfo.hasNextPage value (null, 0, "", [], or {}) is coerced by truthiness. An empty malformed page is reported complete: true, source status: ok, with a successful-refresh timestamp.
2. A Content-Length conversion failure in WandbPages.fetch_snapshot_page leaves its newly opened response unclosed because ownership has not yet reached the collector's finally block.

A three-file candidate, overwatch-wandb-response-guards.patch, modifies the provider and adds tests/test_kev_laya_wandb_response_guards.py and docs/WANDB_RESPONSE_GUARDS.md. Patch SHA-256: 8c2c12a6804f58ccba3880ede554604429678cdf062147e9744f61167946b41a. The candidate requires an actual Boolean and valid continuing-page cursor before accepting that page's records, retains earlier valid pages on failure, and closes an opened response on header-extraction/conversion failure. Valid-page behavior, payload accounting, registry identity, existing length interpretation, bounded requests and cooperative-deadline limitations are preserved. It does not claim transactional validation of every other field in a page.

Measured here: the same 66 synthetic source tests give 40 failed / 26 passed / zero skipped on the unchanged provider, and 66 passed / zero failed / zero skipped on the candidate. Tests use the actual hash-matched provider and standalone telemetry validator, injected in-memory pages/opener, and forbidden socket connections. Compilation, patch apply/reverse checks, three exact postimage comparisons and overlapping-edit refusal passed in disposable partial source copies. Python was 3.13.5, below Overwatch's required minimum; the unchanged preflight exits 2, unsupported_python_requires_3_13_12_to_before_3_14. Ruff was unavailable, including an unsuccessful offline recovery attempt. These are not full supported backend, live transport, cache/report/UI, or native evidence.

The candidate is not committed, SSH-signed, pushed, hosted-validated or merged. It is supplied with source preimages, tests, patch hashes and before/after receipts in the accompanying chat bundle. The configured workstation signer and primary checkout/service are not reachable from this execution runtime; no differently signed API commit was substituted. Integrate only into the existing PR after inspecting the primary checkout and preserving unrelated edits. Do not reapply the earlier cache/CI patches already on PR #1.

The native blocked-by relationship to #2 is unchanged. Its current hosted backend job 108581702145, run 36305664862, was read directly: Python 3.13.12 setup succeeds, then missing_authorized_tscore_index_token exits 2 before locked installation/lint/tests. Existing approved private-index authentication and bounded authorized S3/W&B runtime targets remain necessary; no secret was obtained or configured, remote exchange performed, run identity changed, or WSL operation issued. Related producer gate: Layev #10. Live acceptance remains unverified.

Source-repair delivery gate

• Review and integrate the response guards with the configured verified SSH signer into PR Integrate Layev model runs into Overwatch #1; run supported locked backend/lint and required hosted checks. Source tests alone do not satisfy the live criteria below.

Gap and evidence

Overwatch's current provider implements bounded S3 and explicit W&B GraphQL reads, but the integration validation record says production W&B/S3 exchanges were not performed; transport tests use instrumented doubles. The production W&B endpoint/query and real SDK/network behavior remain unverified. This is separate from the local collector-to-UI gate in #3.

Acceptance

• With existing authorized non-production S3 and W&B sources, run bounded end-to-end collection against the actual provider endpoints without committing credentials or request content.
• Record payload bytes, request/page/read counts, timeouts, partial-source status and preserved prior cache after budget exhaustion, outage and malformed remote responses; verify no hidden eager pagination or unbounded payload read.
• Verify actual W&B GraphQL response shape and S3 body/ContentLength handling, including redirects/retries behavior. State honestly that the cooperative deadline does not cancel in-flight SDK/credential work.
• Confirm source registration and last-success freshness remain distinct from collection completeness. Keep any credential-dependent gate open with its exact blocker.