Rank the noun-pass before slicing instead of first-N in document order

[Samhit21]

Context

PR #41 added a deterministic noun-phrase pass to the JD extractor. The output is capped via allNouns.slice(0, NOUN_PASS_CAP) at src/lib/jd-match/extract-jd-terms.ts:210 — first-N in document order,
no ranking. The cap and the silenced overflow round-trip honestly to the UI ("the noun-phrase pass keeps the first ones it finds and drops the rest" in src/components/features/JdMatch.tsx:72-77), but
on a JD that opens with a marketing paragraph the first 25 capitalized phrases tend to be company-name fluff and the actually-informative phrases get dropped.

Goal

Rank noun-pass hits before slicing so the surfaced 25 are the most informative ones — and update the UI copy back to match.

Suggested approach

Deterministic ranking only — no LLM. A few cheap signals to combine into a score per hit:

• Hit count in the JD body (more occurrences → more central).
• Hit count in the requirements portion of the JD (heuristically: lines in a section whose heading matches /requirements|qualifications|you'll have|what you bring/i).
• Presence in the existing weak-filler stoplist (NOUN_STOP_PHRASES) at src/lib/jd-match/extract-jd-terms.ts:116-120 — extend it with company names, city names, and common product names as needed rather than introducing a parallel list. Keep additions small and inline-documented.

Sort by score, then slice. nounsDropped continues to record the overflow.

Acceptance criteria

• Noun-pass output is ranked before the NOUN_PASS_CAP slice
• Existing tests still pass; new test fixture with a marketing-heavy opener confirms ranking promotes the informative phrases past the cap
• JdMatch.tsx copy updated to honestly describe the ranking step (per the team's copy-claims-round-trip-to-code rule)
• NOUN_PASS_CAP docstring at src/lib/jd-match/extract-jd-terms.ts:106-115 updated to match

Out of scope

• LLM-assisted ranking (separate milestone)
• Industry- or role-specific weights

[s-annam]

Clarification (rewrite) — Body corrected to point the stoplist work at the existing NOUN_STOP_PHRASES weak-filler list in extract-jd-terms.ts:116-120 (reuse gate: extend, don't build parallel) and to refresh three drifted line references (:213 → :210, :105-114 → :106-115, :73-77 → :72-77); issue also moved M1 → M3 · JD Matching as this is JD-match relevance work, not parser hardening.

[s-annam]

Implementation Complete — 845ccd1

Acceptance Criteria

• Noun-pass output ranked before the NOUN_PASS_CAP slice — rankNounHits() scores each hit by body frequency + 2× requirements-portion frequency, stable-sorts (ties keep document order), then slices.
• Existing tests pass; new marketing-opener fixture confirms ranking promotes informative phrases past the cap — new test in extract-jd-terms.test.ts (30-phrase fluff opener + two recurring informative phrases inside a Requirements block); both survive the cap, nounsDropped > 0. Full suite 454/454.
• JdMatch.tsx copy updated to honestly describe ranking — footnote now says hits are ranked by recurrence weighting the requirements section, not "keeps the first ones it finds."
• NOUN_PASS_CAP docstring updated — now describes the ranking step.

Additional Changes

• NOUN_STOP_PHRASES extended with company/marketing-fluff fillers (filtered upstream in extractNounPass, so no double-filter in the ranker).
• Deviation: city/product names were not enumerated in the stoplist (false-positive risk vs. real product-named skills); frequency ranking already demotes one-off proper nouns.