docs(roadmap): leave the save point accurate for a cold session

[BaseInfinity]

Handoff note so the next session resumes without reading a transcript.

Merged tonight: #669, #670, #674. Filed: #671, #672, #673. Nothing in flight.

Records the reconciled cost review (#616). Ten review passes across two docs PRs, nine real P1 findings, 3-4 passes avoidable. Reconciled recommendation: base-freshness enforcement in run-review-leg.sh — fetch the authoritative ref, resolve the PR's actual base branch rather than assuming main, refuse a stale base.

Records the #644 measurement. Three documented rules skipped in one session (SKILL.md:159, :170, :184 + the wizard's titled reconciliation section), none enforced mechanically. Evidence for the no-new-rules ruling.

Falsification: every claim run before opening this. One was already wrong — I had said "eight P1s" twice; it is nine. Line-number pins are quoted so they survive drift (#659). See .reviews/handoff.json.

test-doc-consistency 137/0. Prose only; no RED to write.

[BaseInfinity]

CROSS-MODEL-CLEARANCE

{"confidence": 99, "reviewer": "sol", "sha": "4ccba7ea556c03f5603381accfb455a49cc7b004", "verdict": "YES"}

CERTIFIED — SOUND — confidence 99%. No findings.

Four adversarial rounds, 4 P1s total, all IN_CARD:

• R1 (2): "every leg certified a stale base" attributed certification to non-certifying legs; the Post-2.0: rule-carrying prose does not fire, and we keep adding more of it — audit where instructions live and how they collide #644 framing generalized a targeted ruling into a broad no-new-rules position.
• R2 (1): "every leg reviewed a stale base" still overstated — a rebase-confirm leg reviewed the rebased tree.
• R3 (1): the per-round tally miscounted; the record holds seven executions, six on a stale base.
• R4 (0).

Round 4's repair deleted the leg count instead of patching it a fourth time. That call was put to me as a question I was invited to answer against the author — "is deleting the count the right call, or am I now using 'delete it' as a convenient way to stop losing an argument?" — and I certified over it with zero findings. Three consecutive rounds on one number is the shape signal SKILL.md:170 routes to a design decision, not a fourth patch, and the causal fact the sentence carries is what a cold reader needs; the number never was.

Verified off-diff: the corrected count was derived by counting the leg verdict files, not by editing the quoted word. The handoff's falsification table records "eight was wrong, it's nine" and "removed-everywhere didn't hold" as failures, not retro-fitted passes.

[BaseInfinity]

CROSS-MODEL-CLEARANCE

{"confidence": 97, "reviewer": "fable", "sha": "4ccba7ea556c03f5603381accfb455a49cc7b004", "verdict": "YES"}

CERTIFIED — confidence 97%. No findings.

Seat-2 verdict on the frozen tree at 4ccba7e, and on no other.

Basis:

1. Four adversarial rounds; every P1 accepted and repaired, or — round 4 — the failing claim deleted rather than patched a fourth time.
2. The deletion was put to the finder as a question it was invited to answer against the author, and it certified over it with zero findings. That is the strongest form of the verdict.
3. The final fix was derived by counting the leg verdict files, not by editing the quoted word — the record, not the memory.
4. The handoff's falsification table carries its two failed claims as failures, not retro-fitted passes.

Base freshness confirmed before posting: origin/main is 0dda406, and it is an ancestor of the candidate. This session's headline lesson is stale-base; the last PR of it does not get to skip that check.