fix(supervisor): stop treating execution as verified task success

[jessekemp1]

The supervisor could report turn exhaustion as success, leave evidence-free results unflagged, and feed API completion or heuristic scores back into model routing. Doctor and MCP registration tests also fail on the current baseline.

This change:

• Marks truncation, refusal, empty final responses and tool-turn exhaustion incomplete; preserves partial output and known token usage. Dry runs and unsupported commands cannot count as successful execution.
• Adds attempt identity, unverified/review counts, attempt-specific assessment attachment, and non-overwriting run directories.
• Preserves execution journals while pausing adaptive outcome routing. Heuristic scores remain separate from measured quality; no independent verifier is claimed.
• Repairs sibling-prefix/symlink path escapes and search option parsing.
• Fixes doctor state-directory resolution and tests healthy/missing stores explicitly; uses the public FastMCP client in registration tests.
• Includes missing diagnostic modules in wheels, excludes generated build output, makes wheel checks blocking, and expands lint/offline-test gates.

Validation:

• Guarded, credential-free full CI-scope run: 2,141 passed; 40 skipped; zero failures/errors (JUnit).
• Dedicated integrity tests: 17 passed.
• Touched-file lint, workflow YAML/shell syntax, and diff checks pass.
• Built wheel installed into an isolated target and imported outside the checkout.
• Published Git tree exactly matches the locally tested tree.

Compatibility: success counters still describe execution. Assessments are now serialized as assessment; heuristic quality is heuristic_quality_score, with measured quality_score null. Adaptive routing remains paused until an independent version/attempt-bound verification contract exists.

This is a containment change, not a completed learning system. Transactional event attribution, atomic memory correction, durable task recovery, and installed-machine verification remain open. The test network guard is not an OS sandbox. No merge or deployment is included.