docs: correct the route ladder, refusal wording and verifier count

[DivyamTalwar]

Summary

A conformance pass over published v1.0.0 found four documentation defects. Three of them contradicted artwork or code already shipped in this same repository, which is the part worth fixing before the release is tagged.

What was wrong

The routes table disagreed with its own contract. SKILL.md listed delegate before audit. The decision ladder in references/call-sheet.md and the shipped docs/images/routes.png both use solo → audit → delegate → full → ensemble. The table was the only surface in the repo that disagreed with the contract it defines.

Escalation wording forbade a reduction the contract requires. SKILL.md read "A later call sheet may only escalate… Never silently downgrade", while role-contracts.md instructs downgrading an unsatisfiable ensemble to sequential full. Read together they contradicted each other. Reducing rigour is now described as constrained — it needs the same written evidence an escalation does — rather than forbidden, and the one planned reduction is named.

The CHANGELOG contradicted its own retraction. One entry stated the published config schema "still advertises all three" wire_api values and that the binary is the authority. An earlier entry in the same file had already retracted precisely that: WireApi is a oneOf carrying the single variant responses, so schema and binary agree.

docs/images/proof.png had a caption typo and an outdated claim. It read "the runtimie s own session record", and it predated the mutation warning — implying a refused run leaves the working tree untouched. It does not.

The verifier count is environment-dependent

WALKTHROUGH.md quoted 155 as if fixed. Measured here:

environment	result
PATH=/usr/bin:/bin (what CI asserts)	verify 1.0.0: passed 154, failed 0
Claude CLI on PATH	verify 1.0.0: passed 155, failed 0

The 155th check is the optional claude plugin validate --strict manifest validation. 154 is the invariant floor; the verdict never changes, only the count. Docs now say so.

New artwork

proof.png is regenerated from the launch film's own theme, so README artwork and the video share one source and cannot drift. It now shows the real output: the ERROR: prefix rather than REFUSED:, the specification arriving on stdin (< spec.md — there is no --task flag), and the warning that a write-capable lane may already have modified the tree.

Quickstart smoke test

Four commands, none of which calls a model, spends a token, or writes outside a temp dir. Each was run as written before being documented.

Verification

sh scripts/verify.sh    -> passed 155, failed 0   (Claude CLI present)
sh scripts/verify.sh    -> passed 154, failed 0   (PATH=/usr/bin:/bin)
sh tests/run-tests.sh   -> passed 143, failed 0
sh scripts/cast-call.sh --lane terra --check -> CHECK PASSED, with and without codex on PATH

Risk

Documentation and one image. No script, manifest, lane profile, or test behaviour changes.

Follow-up

This should land before v1.0.0 is tagged, so the release ships docs that agree with the code.