test(eval): D5 C5 prep — warmup cell ceiling + migration map + audit skeleton

[StackOverFlow11]

D5 C5 prep — warmup cell ceiling (C4a LOW-1) + property-test migration map + audit skeleton

Preparation for the D5 C5 four-leg audit, three pieces:

1. Harness hardening (35de346): the bounded warmup_left_extension class gains a cell ceiling (lookback_depth−1) × |frozen symbols|. This is a structural bound, not a calibrated one — every cell the classifier can place in the class is a distinct (date, symbol) pair inside the warmup boundary, so an overshoot means the boundary/counting logic itself is broken (fails closed). Amplitude inside the region stays unbounded per the standing C4a ruling (the two loading geometries legitimately differ exactly there; outside the boundary everything stays gated at 1e-12). Mutation evidence: removing the ceiling conjunction turns exactly the ceiling test red.
2. Property-test migration map (docs/factors/d5_property_test_migration_map.md): all 59 tests across the 10 legacy runner test files mapped to the new suite — 7 moved / 50 covered / 2 carrier-retirements with reason R1 (the retired extract_metrics superset keys are all still produced by analytics/eval/standard.py into section payloads; only the convenience carrier is lost). Coverage strictly non-decreasing; C6's R16 audit consumes this table directly. PR-M having zero runner tests (catalogue BUG 3) is recorded, with the new suite's补齐 tests counted NET-NEW.
3. C5 audit report skeleton (docs/factors/d5_c5_audit_report.md): method (four legs, 77/77 hard gate, frozen baselines read-only), the consolidated 14-entry pre-registered drift table (catalogue §七…§七之五 + handoff §3), and empty per-factor / with_book-decomposition / verdict tables to be filled by the audit run.

Review

Adversarial (separate executor): APPROVE-WITH-NITS — ceiling derivation independently re-proved (incl. the two directions the implementer didn't test: one-cell overshoot FAILs, exact-ceiling passes); a live pooled-abuse probe (2,700 cells of 1.0→999.0, monotone monthly shape, passes — registered as NIT-2: the known pooled early-region amplitude exposure, ~7% of the grid, already adjudicated in C4a and logged in audit §六.1); 15 node-IDs spot-checked (incl. all 7 "moved"); 5 drift-table entries re-checked against code constants. Two NITs; NIT-1 (docstring swallowed the (date,symbol)-uniqueness premise of the "by construction" argument) fixed post-review in 39e7f9c with the reviewer-prescribed wording — docs-only, tests unchanged; NIT-2 kept as a registered observation for the C5 verdict's reading discipline.

Gates (independently rerun by the lead)

pytest 2537 passed (node-ID diff +4/−0), ruff clean, run-phase0 anchor ic 0.9600 / annual 0.8408, secret scan 0. The full-audit background run's progress is unaffected by this branch (separate worktree).