docs: FIELD-NOTES.md — the failure mode behind every gate

[MyAlterLego]

The gates in #6 each exist because an analysis shipped a specific error, but that reasoning lived only in a local SKILL.md — not in this repo, and on an ephemeral filesystem. The checks would have outlived the explanation, and the first person to hit a red gate would have no way to judge whether it was telling them something true.

Seven failure modes, each with the evidence and the gate that now catches it:

1. Four false absence claims — asserting a control was absent without reading the code implementing it. One rated band-1 R0 and recorded as CONFIRMED. Same mechanism every time: read the use, infer the default, never read the declaration or the mount. The bias runs one way, so it always inflates severity.
2. 2,508 guard candidates the analysis never consumed — including a WWW-Authenticate header inside the middleware a finding declared did not exist.
3. A belief credited as a working control because sourcedFrom named a hop, not an origin — invalidating six remediations that were placebos against a forged token.
4. Two generic RPC bridges that survived two passes and an adversarial review, because a :param route collapses into one REST control action.
5. 0 upgrades in 96 verdicts — a property of the brief, not evidence of completeness.
6. A verdict filed but reverted by the apply step, while the scorecard claimed it had been checked.
7. A hand-typed count that drifted from the grid.

It also records the three false positives these gates produced on first run, because a gate that cries wolf gets disabled — that is a failure of the gate, not a cosmetic issue. The generalisable one: the evidence standard must match what the analysis was derived from. A codebase analysis cites code; a design-document analysis cites the document.

Docs only — no tool changes.