feat(web): semantic extraction on Firecrawl + dashboard semantic-vs-positional hit-rate

[QodeXcli]

Two follow-ups to the semantic web-extraction PR (#90): make it work across all backends, and measure its value.

1. Semantic selection on the Firecrawl backend

Firecrawl content-mode returned scraped markdown capped by a blind md.slice(0, 1500) head cut — the same positional weakness. Now:

• parse.ts gains capMarkdown(md, query, cap) (PURE) that trims with the same selectRelevantPassages — semantic when the search query is given, head+tail otherwise
• mapFirecrawlResults threads the search query + an onExtract callback, and strips base64 image bombs
• Backward compatible: no query → head+tail window, still ≤ cap (verified the existing ≤1500 assertion holds)

So a Firecrawl content-mode search now returns each result trimmed to the passages most relevant to what you searched for.

2. Dashboard hit-rate counter (show the real value)

• extract-metrics.ts — a PURE aggregator over a best-effort append-only JSONL under ~/.qodex (always on, deliberately not the opt-in telemetry DB). parseExtractMetrics → {semantic, headTail, truncated, semanticRate}
• web_fetch + firecrawl record each truncation's mode (whole pages ignored)
• Dashboard panel "Web extract — semantic vs positional": semantic hits, head+tail fallback, and the semantic hit-rate — so you can see that agents pass a query and get the relevant middle instead of a blind window

Tests

extract-metrics.test.ts (+6): aggregator counts/rate/empty; capMarkdown semantic vs head-tail; mapFirecrawlResults semantic + onExtract mode; whole-page ignored. Dashboard asserts the panel + rate render. Full suite 1488 green, tsc clean.