Blizzard and Raider.IO fixtures encode guessed upstream shapes

[Erilla]

Question

Five faults surfaced in the fingerprint feature within hours of it reaching test, and 210 unit tests plus the integration and e2e suites passed through every one of them:

Fault	Fixed in	Why tests missed it
A candidate 404 abandoned the whole sweep	#32	No fixture returned 404 mid-roster; the live measurement saw 23 of 393
Ownership gate excluded never-claimed characters	#34	Fixtures only covered "withheld", never "no account"
discord_profile: "" raised schema drift	#35	Every fixture used a non-empty string or null
403 profile_is_private retried as an outage	#36	No fixture returned 403
Roster class name required but never sent	#38	The fixture invented playable_class: { name: "Mage" }

The pattern is not "insufficient coverage" — coverage was good. Every one of these fixtures encoded a guessed upstream shape, each guess was wrong, and a passing test then certified the guess. The roster case is the clearest: the fixture asserted a field that does not exist in Blizzard's response, so the test could only ever have passed while production could only ever have failed.

To decide:

• Whether the Blizzard and Raider.IO fixtures get rebuilt from recorded live payloads (redacted) rather than hand-written objects, and where those recordings live.
• Whether a periodic check re-fetches a known character and roster and diffs the shape against the fixtures, so upstream drift fails CI instead of production.
• Whether the negative cases each upstream really returns — 403, 404 mid-roster, empty strings, absent optional objects — become required fixtures per gateway rather than ad hoc additions.

A recording harness already exists in spirit: scripts/prototypes/fingerprint-sweep-measurement.ts parsed the live roster correctly precisely because it read only what was there and skipped the rest. It was deleted from the plan's scope as disposable; its leniency was the thing worth keeping.

Related: #30 — nothing here would have been caught without a live run, which is the argument for that smoke test gating promotion.