feat(web_fetch): query-aware semantic extraction — beat Hermes's positional truncate-and-store

[QodeXcli]

Reviewed Hermes's PR #54843 (feat(web_extract): truncate-and-store instead of LLM summarization) and built something semantically better for QodeX.

What Hermes did

They dropped a per-page LLM summarizer (Gemini-class, run on every page >5k chars) in favor of positional truncate-and-store: pages over a 15k char budget return a head+tail window (~75/25) plus a pointer to the full text on disk. Result: 11.7× faster, ~23× cheaper, quality unchanged.

Their tradeoff: positional selection. Their own eval was 3/4 on "answer present in the returned window" — when the relevant passage is in the dropped middle, the agent must make an extra read_file round-trip to recover it. (QodeX's old web_fetch was worse — a plain slice(0, maxChars) head cut, no tail, no store, no recovery.)

What we did — keep their win, beat it on quality

web_fetch gains a query param (what you're looking for). For an oversized page:

• Semantic passage selection — rank passages by the same stemmed TF-cosine we already ship for recall_approach (feat(recall): MMR diversity in recall_approach #78/feat(recall): semantic stemming — paginate↔pagination, authenticate↔authentication #85), keep the title/lede as an anchor, return the most relevant passages in document order. The mid-document answer that costs Hermes a round-trip comes back in the first call.
• Never worse than Hermes — no query → the identical head+tail window (a unit test asserts byte-equality).
• Store-and-recover — full clean text written to ~/.qodex/cache/web/…, with a footer giving the exact read_file path=… offset=… limit=… to page the rest.
• Base64 image token-bombs → [IMAGE: alt] placeholders; real http(s) image links preserved (mirrors their image fix).
• Zero added LLM cost — the ranker is pure lexical math (microseconds), so we keep their speed + cost win and add the relevance they dropped.

The headline test proves it: on a page whose answer sits in the middle, head-tail mode omits it while semantic mode returns it — same budget.

• src/tools/web/extract-select.ts — PURE core, 7 unit tests
• src/tools/web/web-fetch.ts — query param, store-to-disk, recovery footer, base64 handling

Full suite 1474 green, tsc clean.