fix: stop counting eBay's 'Results matching fewer words' padding as sold comps (0.3.1)

[mabry1985]

Why

The first live sold-comp re-price (45 searches, "no errors") produced rows like "54 sold, median $31.43" for a Bretonnian dice set. Opening that query showed eBay's headline "0 results", two "Shop on eBay" filler cards, the divider "Results matching fewer words", then 54 loosely related items (complete boxes at $123, other teams' dice, $5 sprues, $1.34 transfer sheets). RESULT_JS collected every li.s-card on the page. Five unrelated rows sharing p25 = $20.22 and four sharing p25 ≈ $29.7 were the same padded river. The affected inventory file was restored from its pre-run copy.

What

• Page script: finds the divider (leaf element whose text is "Results matching fewer words" / "Results for similar searches"), excludes every card after it from rows, counts them as related_rows_excluded; reads eBay's own headline_count from the count heading; flags query_rewritten on "Showing results for … / Search instead for …". Filler cards were already skipped.
• Tools: ebay_price_check, ebay_search, ebay_price_and_profit and the eBay sources of compare_prices carry headline_count, related_rows_excluded, query_rewritten (only when true) and a notes list: zero exact matches → "do not price from those, broaden"; fewer than 5 next to padding → thin, broaden; otherwise a one-line count; rewritten → every number is for eBay's query.
• Skill: reading the new fields; "the same quartiles for different items is padding, stop and look at the page"; notes are part of the answer.
• Version 0.3.1.

Tests

135 host-free (was 127): zero-match page reports no comps and says why; price-and-profit names the padding; thin match next to padding flagged; healthy page with some padding just says so; clean page has no notes and claims no headline; rewritten query called out; compare carries notes per eBay source and not for Amazon; the page script emits the new keys. Live validation against a padded page and a real one is posted as a comment on this PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_01X1Mou2gxw6Wz9dLCmVocoL

[mabry1985]

Live validation of the new page script (agent-browser 0.27.1, signed-in profile, 2026-09-12):

query (sold view)	eBay headline	count (exact)	related_rows_excluded	0.3.0 would have reported
Blood Bowl Third Season Brionne Barons Bretonnian dice set 7 dice	0 results	0	54	"54 sold, median $31.43"
Blood Bowl Third Season rulebook	22 results	22	51	"17 sold, median $73.96, p75 $123.25" (padded with $123 complete boxes)

The exact-match count equals eBay's own headline on the real query, and the first six rows are all rulebooks ($26.51–$65.00). query_rewritten was false on both, as expected.

Also observed while validating: Chrome 149's built-in Gemini side panel ("glic") can open in the headed window and agent-browser adopts it as the active page, after which navigations land there and the tools report "could not find the results list … The browser ended up at: https://gemini.google.com/glic". Not addressed in this PR — the error already names where the browser went; closing that panel fixes it. Follow-up: have Browser.open() verify the landed host and fall back to a fresh tab.

[protoreview]

QA panel review — PASS

code-review-structural · head ffa604cae3b1 · formal

All four LLM finders returned zero findings and the structural pass hard-stopped with no output, so the panel produced no findings to verify. The verifier confirmed the empty array. No prior requests to disposition. The one thing to flag: the structural pass's hard-stop means the panel's coverage is thinner than usual — if this PR touches non-trivial production logic, the silence is a gap, not a clean bill. No panel disagreement (unanimous zero). No verification changes (empty in, empty out).

No findings — the review came back clean.

findings JSON (machine-readable)

[]

[protoreview]

Promoting the PASS verdict for head ffa604cae3b1: all checks terminal-green, zero unresolved review threads. (approve-on-green)