fix: drop the >80%-broken-locators full re-research retry

[DavertMik]

Problem

Trace d9fd5d5a14aba27bc86cf925f57dc8a1 spent 67s opening the New Step page, because the researcher researched the same URL twice — a researcher: /steps/new span containing a nested researcher span for the identical URL against the identical DOM:

researcher: /steps/new                       67s
  ├── invoke_agent  28s   ← full research report (58 locators, 4 containers)
  └── researcher: /steps/new                 35s   ← re-research, SAME url, SAME DOM
        ├── invoke_agent  21s   ← report regenerated from scratch
        └── invoke_agent  13s   ← fixBrokenSections

Two of the four containers were imprecise: 'nav' matched more than one element (testLocators requires count === 1) and '.detail.detail-view-resizable' matched zero. Those cascade to their child locators, which get marked broken without ever being tested — inflating brokenRatio past 0.8 and tripping the guard that threw away a usable 28s report.

The regenerated report only swapped 'nav' → '.mainnav-menu-body' and '.detail.detail-view-resizable' → '.detail-view-resizable'. That's a container correction — precisely what fixBrokenSections does by continuing the existing conversation instead of regenerating. It ran anyway, 13s later, and is what actually fixed things.

Fix

Delete the brokenRatio > 0.8 retry (8 lines). Broken locators fall through to Stage 3 fixBrokenSections, the mechanism built for them. brokenCount/brokenRatio had no other consumers, and the cascade stops mattering because it was their only input.

Expected: 67s → ~41s on affected pages, with no nested research span.

Trade-off

If a page's DOM is genuinely stale when researched, the deleted 3s sleep + regeneration could recover where fixBrokenSections will now repair against that same stale DOM. The two catastrophe guards — nothing parsed, every container broken — still retry, so only the partially-valid case changes behavior.

Note for the reviewer

The trace doesn't record which retry branch fired, so I inferred it from the 3.243s gap between the first report finishing and the retry starting (3s sleep + locator testing) plus the container evidence. It fits the brokenRatio branch and not the other two, but a debug-log replay would confirm it outright.

Testing

• bun test tests/unit/ — 808 pass, 0 fail
• bun test tests/integration/ — 63 pass, 0 fail
• bun run check:fix — clean

🤖 Generated with Claude Code