fix(SatisfactionCapture): don't fabricate a neutral 5 when there's no real rating

[lewta]

Problem

SatisfactionCapture.hook.ts writes a fabricated rating: 5 whenever it lacks a
real satisfaction signal, in three places:

1. Inference call returns failure → wrote {rating:5,"Inference failed…"}.
2. Inference call throws → wrote {rating:5,"Inference error…"}.
3. Inference succeeds but the model omits/garbles the rating →
   const rating = (… valid …) ? r.rating : 5; quietly substituted 5.

A fabricated 5 is "we don't know" masquerading as "neutral." These rows
accumulate and flatten every today/week/month average — the aggregator
(PAI/statusline-command.sh ~L1422-1438) flat-means .rating and never reads
.confidence, so low-confidence placeholders still count at full weight. In one
real install, 238 of 881 signals (27%) were "Inference failed" placeholders.

Fix

Introduce isValidRating(x) as the single source of truth (a real number in
1-10). Persist a rating only when it's valid; in every other case log for
observability and write nothing. The success path no longer coerces a
missing rating to 5. No behavior change when the model returns a valid rating.

Alternatives considered (and why skipping wins)

• Lower the confidence on the placeholder row — no effect. The aggregator
  flat-means .rating and ignores .confidence; the failure rows are already
  written at confidence: 0.3 and still skewed every average.
• Write a sentinel rating excluded from averages — would require teaching
  every rating consumer (the statusline jq, LearningPatternSynthesis.ts, …) to
  recognize and exclude it: a larger, multi-file change for no added benefit.

Skipping keeps the change small while preserving observability — failures are
still logged to stderr. If failures should be queryable, the right home is a
separate failure stream, not the ratings signal.

Testing evidence

Two tests under Releases/v5.0.0/.claude/hooks/tests/:

• isValidRating.test.ts — unit test of the validity guard (accepts 1-10;
  rejects null/undefined/0/11/NaN/Infinity/strings/objects). bun test → 2 pass.
• SatisfactionCapture.failure-path.test.sh — behavioral: forces a deterministic
  inference failure (removes claude from PATH), runs the hook against a
  throwaway temp PAI_DIR (never touches a real ~/.claude), asserts no row
  written. Before/after on identical input:

[PRE-FIX ] inference failure → wrote 1 row {"rating":5,"sentiment_summary":"Inference failed: …"}
[POST-FIX] inference failure → wrote 0 rows; stderr: "… skipping signal (no rating written)"
→ PASS (exit 0)