feat(introspection): track refusal-recovery rate, not just refusal count (Closes #1366)

[Lexus2016]

Closes #1366 — Child B of #1327.

The gap

Child A (#1356, merged bb0a325cc) added RECOVERY_BEFORE_REFUSAL_GUIDANCE: before refusing for a missing capability, propose an alternative path.

Nothing measured whether it worked. _REFUSAL_RE matches these identically:

• "I can't do that — there is no tool for it." ← the failure [UX] Soft-capability refusals plateau at 51/week — agent declines instead of proposing alternatives #1327 is about
• "I can't call that API directly, but I can query it with a script instead." ← exactly what [UX] Soft-capability refusals — Child A: Recovery-before-refusal prompt nudge #1356 asks for

Both bumped one counter, so the 51/week signal could not move even if the guidance were perfect — and a realized-impact verdict on #1356 would have been forced to no-signal for lack of a comparable rate.

Change

_RECOVERY_RE — a turn counts as recovered when it refuses AND offers a concrete alternative. Deliberately conservative: pivot phrases that introduce a different path (instead, alternatively, workaround, however I can, what I can do), not mere hedging. Over-matching would inflate the rate and make #1356 look effective when it is not — worse than under-counting.

Emitted additively, per session and in the digest:

Field	Meaning
refusals_with_recovery	count of refusals that offered a path forward
refusal_recovery_rate	recovered ÷ refusals

The denominator is refusals, not sessions — this measures the quality of the refusals that happened, not how often the agent refuses. Both numbers move independently and both matter.

refusals / refusals_or_access_denied keep their exact current meaning; a back-compat test asserts the aggregate is unchanged.

Measured on the real corpus

sessions_scanned: 4629
refusals:         143
with_recovery:     51
recovery_rate:   0.3566

So #1356 currently fires on roughly a third of refusals — two thirds are still bare. That is the baseline Child C (recovery dispatcher, ~150 lines, still unfiled) needs in order to demonstrate it improved anything.

Success criteria

• A refusal that proposes an alternative is counted as recovered
• A bare refusal is not
• refusal_recovery_rate appears in the digest
• Existing refusals counts are unchanged

43 passing in test_introspection_extract.py (9 new), 82 across the introspection and realized-impact suites; ruff clean.