feat(plan): full autonomy by default, and retire the stale model priors

Three changes, one judgment surface, interleaved in the same files:

AUTONOMY. The chain-wide question budget defaults to ZERO at every scale:
infer, state each inference in one line the user can correct in passing,
proceed to the plan. Scope, ambition, and technical depth are never
question material. Budget above zero comes only from the person —
round((1 − autonomy) × 6) via the /plan-tune profile, which is now
actually consumed (its observational-only disclaimer was true and is
gone). Goal question inferred, ambiguous → Builder; Builder's five
sharpening prompts became commitments; independent questions batch four
per round; consent gates with nothing to bundle into skip-with-note.
Aligned across all seven budget stamps.

HUMAN-ERA ANCHORS. The eng complexity gate challenges unjustified
STRUCTURE, never diff size; ceo's >15-files→REDUCTION trigger is gone;
autoplan auto-approves by blast-radius + no-new-infra + no-authority-
change instead of a 5-file proxy; spec answers its own discovery
questions and DECIDES design choices instead of interviewing; issues
sized by review-ability, not human build-days; hardcoded human/CC speed
ratios removed as drift-prone.

CAPABILITY PRIORS. A dated, sourced rule in all five plan specialists:
never justify narrowing scope by asserting an LLM can't do something.
2026 evidence: pairwise matching 92-99 F1 with the error direction
INVERTED vs rule engines (3% vs 14% false merges), objective judging at
kappa 0.87, grounded hallucination low single digits. What still fails is
named (subjective preference judging, mega-schemas, small-model
clustering). Uncertain load-bearing capability → keep the ambitious
design, plan a 30-minute spike with ~50 labeled examples and an 85-90%
agreement gate. Measure it this afternoon; don't downgrade the product.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: 2 people