fix(skills): make honesty mechanics falsifiable and route failed batches

The abandonment rule rested on "if the force-fit still feels strained" —
a test the operator grades themselves, and one a model that wants to look
clever never fails. Replace it with the redundancy test the worked example
already articulated: keep a force-fit only if the idea was unreachable
without the stimulus.

random-stimulus:
- Redundancy test replaces "feels strained"; abandonment expectation
  raised 1-in-5 -> 1-in-4 to match a rule that actually bites.
- Separate the batch claim ("the third or fourth stimulus pays off") from
  the per-stimulus rule ("stop grinding one dead object"). These were
  conflated in one sentence and contradicted each other.
- Step 2 minimums are now countable: >=5 categories, >=2 concrete,
  >=1 abstraction, no adjacent same-pool draws.
- Kind is a property of the stimulus, not its pool — hourglass and
  circadian rhythm share Time & Cycles. Step 3 labels each chain
  `Pool [concrete|abstract]` so a reader can check all three minimums.
- Step 5 scans the abandonments too: a shared reason for failure is a
  finding, not noise.
- A failed batch no longer answers itself by drawing more stimuli.

Failure hand-offs (random-stimulus, provocation, analogy, worst-idea):
each names the diagnosis and suggests the fitting technique as the user's
next move, explicitly not running it — the router permits one technique
per pass.

References:
- worked-example: chains labelled with pool and kind; category summary
  corrected (six categories, not eight — river delta is Geography, not
  Water). Transcript reasoning untouched.
- stimulus-pools: five accidental duplicates removed, three intentional
  dual-pool entries disambiguated, `erosion` added (the example drew it
  but no pool contained it). 620 stimuli, zero duplicates.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>