chore: harden verdict panel= parsing + fix stale README test count (137 vs 450)

[jakecelentano]

Scope

Two small hardening/hygiene items in the review poller area:

1. panel= parse fragility. _PANEL_RE (lib/tms/review_poll.py:206, panel=(.+?)(?=\s+\w+=|$)) assumes the panel value is terminal or followed by another key= token. A verdict line whose panel value is followed by a bare non-key= word (or contains stray spaces) over-captures into the panel field. Replace with a tokenizer or tighten the lookahead, with regression tests for malformed-but-plausible verdict lines.

2. README test-count drift. README.md:161 claims "137 tests, ~0.5s"; the suite collects 450. Update, or better, drop the hardcoded count so it cannot drift again.

Acceptance criteria

• Malformed panel values covered by regression tests (trailing word, embedded spaces, empty panel)
• parse_verdict_line behavior on those inputs documented in the docstring
• README test claim corrected or made count-free

References

• Found during the 2026-07-20 stack evaluation
• lib/tms/review_poll.py:64-129,200-238, README.md:161
• Related: feat: tms verdict parser + capture support specialist slots (reviewer-security(minimax-m3)) #107 (specialist slots in the same parser — coordinate)