fix: defer failed skill assessments

[theonly1me]

What changed

• Send a Claude-compatible JSON schema for skill assessment while keeping response limits in local validation.
• Defer remaining skill assessments after an engine failure and leave them eligible for the next run.
• Report deferred skills from learn --deep and cover schema compatibility, retry state, malformed results, and CLI output.

Why

learn --deep --apply could finish profile reconciliation and then fail skill assessment because the generated schema contained constraints Claude rejects. Engine failures also escaped the existing deferred work path and ended the command with Skill assessment engine failed.

How to verify

bun run typecheck
bun run lint
bun test

Claude Code 2.1.236 completed a funded synthetic assessment and deferred an assessment with an intentionally small budget.

Data handling

• Skill text continues through resolveRedacted before the learning engine call.
• Provider errors are not logged, persisted, or added to test fixtures.
• Tests exercise learn and the skill maintenance entry point with synthetic data.