docs,test(dogfood): record repair evidence and stabilize installer CI

[Endofthestars]

What changed

• add the registered Round 2 analysis and implementation contracts
• preserve Round 1 Protocol and Experiment traces for comparison
• add restart-before and restart-after canonical projections
• add Deep Doctor, Chronicle verification, and SHA-256 evidence
• document the repair path from PT-002 through the sealed REPAIR Decision
• make installer host-detection mocks command-specific in clean CI environments
• normalize both sides of macOS symlink path assertions

Why

The first dogfood trial could not complete because its recorded Run IDs and
replay_success metric did not match frozen PT-001, which required
RUN-000 and metric score. Athanor correctly rejected the inconsistent
pilot completion.

Round 2 preserves that failed trial and records a new Protocol-aligned path
using PT-002, EX-002, and RUN-101 through RUN-105.

The first PR run also exposed pre-existing installer test assumptions: a broad
shutil.which mock made an absent Codex CLI appear installed, while macOS
reported equivalent temporary paths through /var and /private/var.

Impact

This adds evidence under examples/phase2-final plus test-only portability
fixes. It does not modify runtime behavior or canonical .benchwork/ state.
The evidence shows that the repaired cycle completed analysis, assessment, and
a user-confirmed REPAIR Decision while retaining the failed excluded Run.

Validation

• SHA-256 verification passed for all exported evidence
• restart-before and restart-after snapshots are byte-identical
• Deep Doctor passed with 53 Chronicle events and 37 replayable objects
• Alembic recomputation matched RB-001 and its registered Sigil
• python3 -m unittest tests.mcp.test_runtime.MCPRuntimeTest.test_complete_repair_cycle_uses_only_control_plane_tools
• affected installer modules: 26 tests passed
• local independent review completed with no remaining findings