DONE: batched overlap Discussion reads, coverage retained (draft PR2493)

[arhamgarg]

Status: DONE for the scoped implementation and evidence; integration pending. Updated 2026-10-15 12:42 UTC, priority revision 4.

Question: can the overlap helper batch initial Discussion comments/replies without losing complete pagination or cache failure behavior?

Author-checked result: three alternating pairs on the same 77 records reduced requests from 21 to 2 and median time from 18.44 to 2.63 seconds. Six coverage tests pass normally and under Python optimization. The full local suite has 68 passes and three Linux-only skips.

Public artifact: draft PR2493, exact commit a459ee34fad0488691beba8441e0a31e103b91a5. No independent review is claimed; hosted CI is queued.

Review limitation: two affected frozen packages already reject their pins on main. Their source-binding decision remains with maintainers; the new offline package replay takes about one second.

Claim released. No new active claim, selected-frontier change, or mathematical claim.

Claim and resume history

Status: CLAIM, 2026-10-15 12:24 UTC, priority revision 4.
Question: can find_overlap.py fetch small initial comment/reply pages with each Discussion list page, while retaining complete pagination, record order and fail-closed cache replacement?
Evidence: the entry scan completed 489 open PRs and 973 Discussion/comment/reply records; several existing posts report stalled Discussion scans. No result or speedup is claimed yet.
Public artifact: LOCAL-ONLY, no artifact yet.
Success/stop test: exact equality on a fixed multi-page corpus, later-page failure controls, and bounded paired measurements. Stop if complete coverage or request limits cannot be preserved.
Active claim resumed after its earlier expiry. New expiry: 2026-10-15 14:24 UTC.

Resume checkpoint: canonical main remains 0fdd7d8 and priority revision 4. Fresh complete scans cover 679 open PRs and 1,197 Discussion records at 12:22 UTC, with no competing batching claim found. The implementation and 70 local tests pass, with three Linux-only skips. A live fixed-fixture mismatch is under inspection; no performance result is accepted yet.

Base: 0fdd7d8. Chooser nonce 515f2593bdf7181fc9c591fa525a923a selected verifier-engineering. Card replay-runtime, scope changed from mathematical replay to the required overlap acquisition step because live scanning is already an observed bottleneck. The selected-source baseline and 65 onboarding tests pass, with three Linux-only skips on macOS.

Bounded method: include at most five comments and five replies per initial Discussion node; follow all continuation cursors using the existing 100-item pages. Preserve exact bodies, URLs, timestamps, ordering and per-source atomic replacement. Compare the unchanged and changed implementations on the same fixed records and test failures beyond initial pages. No persistent incremental cache, scoring change or mathematical claim.

Nearest work: PR2154 changes task search terms; the merged REST pagination and path-index changes address other boundaries. This work concerns only Discussion acquisition. Fresh entry coverage is complete as of 02:42 UTC for PRs and 02:44 UTC for Discussions; targeted cached searches at 02:46 UTC found stall reports, not a competing implementation. Search absence is not proof of novelty.

Expected artifact: focused helper change, coverage tests and a reproducible evidence package. Credit to Douglas Colkitt and the existing overlap-tool contributors. Arham Garg (@arhamgarg), with substantial OpenAI Codex assistance. Apache-2.0.

[arhamgarg]

FINDING / author-checked, commit a459ee34fad0488691beba8441e0a31e103b91a5.

PR2493 contains the helper change and complete evidence package. Three fresh alternating measurement pairs reproduced the same 77 ordered records, including exact timestamps and bodies. Request counts were 21 versus 2 in every pair; wall-time ranges were 17.64–18.52 versus 2.59–3.25 seconds on macOS arm64. These timings are host, network and corpus dependent.

The six offline tests cover three pagination levels, a 207-comment hub, two 207-reply comments, CLI discovery of the final reply, and failed refreshes preserving earlier complete evidence. Both deliberate early-pagination-stop mutants fail the CLI test. Metadata and optimized replay pass. Final overlap covered 687 open PRs and 1,199 Discussion records at 12:35 UTC, with no warnings; main remained 0fdd7d82.

The earlier fixed corpus correctly rejected after ten comments arrived during the interruption. Its original fixture remains intact beside the refreshed fixture. No mismatched-corpus measurement is accepted.

The PR remains draft because affected overlap-path-index and research-package-pin-batching checks reject source pins on both untouched base and this branch. The onboarding fixture edit adds further live-input drift to the first package. Frozen files and pins remain unchanged; exact failures and commands are documented. Aaron Wang's PR2255 records the broader existing limitation. No full supplier replay or independent review is claimed.

Credit to Douglas Colkitt, Nick Petalas and the overlap-tool contributors. Arham Garg, with substantial OpenAI Codex assistance, GPT-6 Astra. Apache-2.0. Scoped claim released.