Record the assurance baseline at 40c521b in the roadmap

[404SecNotFound]

Summary

A reconciliation before any assurance change. Docs only, one file (docs/ROADMAP.md). Nothing in src/ or reference/ was modified to produce these results.

It adds Phase 9, 9.0, a dated baseline with five findings. Each was reproduced on this tree with synthetic inputs. The probes loaded the shipping worker as-is with a stand-in self, and called the shipping split function directly, rather than reimplementing either.

#	Finding	Status
9.1	The inheritance plan ignores "The strips need the password too". In Chromium, the plan still says any 2 of 3 shares open the backup on their own.	Still present
9.2	Verify-only and rehearsal receive the plaintext in the page, which zeroes it without rendering it.	Still present, mitigated
9.3	Creating a backup derives the password once per way in: 1, 2 and 3 times, for PBKDF2 and Argon2id alike. Password AND shares already derives once.	Still present
9.4	Shamir polynomial coefficients are still non-zero after the split returns. Caller-supplied coefficients are untouched.	Still present
9.5	Cold recovery from the four kit files with networking removed. It fails without wheels, recovers two containers byte for byte with the pre-downloaded wheels RECOVERY.md describes, and refuses a tampered file with no output.	Works as documented

It also marks owner item O1 as done, because the description and 16 topics are set.

Status: 9.0 verified. 9.1 to 9.4 not_started, one PR each, based on main, in the order 9.1, 9.3, 9.2, 9.4.

Checks on this tree

All exit 0 on 40c521b:

• typecheck, test:readme, test:calibration, test:release-gate, test:reproduced-manifest, build
• test:keymaker 297 passed. test:keym2-dispatch 76 passed.
• test:fuzz2 6,023 assertions. test:fuzz3 4,853 assertions. test:shamir 19 assertions. test:secret-erase-core passed.
• keym2.py selftest 680 checks. crosstest2.py 445 passed, v2, v3 and v4 byte for byte. recovery_test.py 185 passed.

Not run

• Firefox and WebKit are not installed in this environment, so their coverage is this PR's CI run.
• No physical authenticator was available, so nothing here covers real passkeys.
• No negative control applies, because this PR adds no test.