fix(pulse): map adapter models to inference levels that exist

[elhoim]

The problem

AdapterRunner.ts's modelToLevel() could only return "fast", "smart" or "standard". Those names were removed from the inference layer on 2026-06-10. Inference.ts now validates against low | medium | high | max and throws on anything else:

[Inference] unknown level 'smart' — use low | medium | high | max

So every level the adapter could produce was rejected. Nothing caught it: callInference and runAdapter have no catch, AdapterCli.ts top-level-awaits runAdapter so the process dies, and RebuildAll.ts uses Promise.allSettled, printing each rejection as an unlabelled "unhandled" row. Only the cache-hit, no-sources and missing-prompt early returns avoid inference, so 100% of actual adapter builds failed.

Why it compiled

This is the part worth fixing properly. Inference.ts declared its own copy of the level names:

export type InferenceLevel = 'low' | 'medium' | 'high' | 'max';   // a copy

models.ts already had EffortLevel as the real source of truth. Two hand-maintained lists of the same names means a rename in one leaves the other stale and still type-checking — which is exactly what happened. The type mismatch didn't compile-error because the adapter was typed against a copy that agreed with itself.

The fix

Bind both ends to one source.

export type InferenceLevel = EffortLevel;                              // alias, not a copy
const VALID_LEVELS: readonly InferenceLevel[] = Object.keys(EFFORT_MODEL) as InferenceLevel[];

The type now IS models.ts's level type, and the runtime validator derives from EFFORT_MODEL's keys. Rename a level in models.ts and every stale call site fails at the type level rather than at runtime. The error message and the CLI's own validation also read from VALID_LEVELS instead of repeating the literals a third and fourth time.

Mapping, extracted to adapters/model-level.ts and typed InferenceLevel:

model family	level
haiku	low
sonnet	medium
opus	high
fable	max
anything else	medium

Cheap models map to cheap tiers, and an unrecognised model gets a sane middle default rather than the most expensive one.

Failures stop being unlabelled. callInference catches a rejection and returns the same structured failure shape as every other error path, so one bad page reports as a named page with a real message and the batch keeps going.

Tests

test/pulse/adapters/model-level.test.ts, 8 cases:

8 pass
0 fail
49 expect() calls

Separately verified that every level the mapping can emit is one the inference layer actually accepts, checked against EFFORT_MODEL's real keys rather than a hardcoded list:

VALID_LEVELS (from EFFORT_MODEL): max, high, medium, low
✓ claude-haiku-4-5       → low
✓ claude-sonnet-5        → medium
✓ claude-opus-5          → high
✓ claude-fable-5         → max
✓ gpt-5                  → medium
✓ <empty>                → medium
✓ weird-unknown-model    → medium

No API calls are made by the tests or the check.