auto-research page: mirror 20 Aug-2026 works + fix 2 rows stranded in thead

[ldkong1205]

Mirrors the Aug-2026 auto-research update into the project page's S1–S8 stage tables. Companion artifacts: paper f9692ed (Overleaf), README PR #17.

Added (20 rows)

Stage	tbody rows	+
S1 Idea Generation	38 → 43	5
S2 Literature Review	58 → 65	7
S3 Coding & Experiments	45 → 51	6
S6 Peer Review	48 → 50	2

Stage counters bumped by the same deltas. Source of truth is the paper's appendix_s1–s8, so these mirror the appendix additions rather than the README's (the appendix keeps analysis and position papers prose-only).

Nine new rows carry v-conf venue pills — EMNLP'26 ×4, ACL'26, CIKM'26, MICCAI'26 — rather than v-arxiv, matching venues confirmed in arXiv metadata.

Pre-existing bug fixed: two S1 rows were stranded in <thead>

S1's first two data rows sat inside <thead> instead of <tbody>. This explained three things that initially looked like separate problems:

• S1's tbody numbering started at 3 and ran to 40, so it was the only stage failing a contiguity check.
• S1 was the only stage whose counter (40) appeared not to match its tbody row count (38). The counter was in fact always correct — it counted all 40 rows; two just lived in the wrong element.
• Those two rows rendered with header styling and were invisible to any tbody-scoped sort or filter.

Both rows are moved into <tbody>. No content changed.

17 of 37 works not mirrored

The page has only S1–S8 stage tables. There is no home for the 7 end-to-end systems (Gemini Co-Scientist, AutoResearch, Eureka, AutoResearch Diagnosis, AutoSR, The AI Engineer, Symposium) or the 7 societal / cross-domain entries, and 3 analysis papers are prose-only by the appendix convention. This is the same limitation recorded in the 260817 pass, where 5 of 23 works had no stage home.

Together with the parallel gap on survey.html (PR #65, 7 of 27 unplaceable), this is now a recurring structural shortfall across the project pages, concentrated in exactly the categories the field is growing into — end-to-end systems and evaluation. Worth considering an E2E panel here and a datasets panel there.

Checks

• Counters equal tbody row counts for all 8 stages; numbering contiguous 1..N in all 8.
• Zero data rows remain in any thead (was 2).
• <tr>/<td>/<tbody>/<thead>/<table> tags all balanced; git diff --check clean.
• Branched from origin/main, so the two unpushed co-author EgoMM commits and the pending yu_yang.jpg deletion in the local clone are untouched.