docs: agent failure modes — validated catalogue + structural mitigations

[Jammy2211]

Summary

Design deliverable for #130 (build-chain campaign PyAutoBuild#155, Phase 5 — deliberately last): docs/agent_failure_modes.md. Docs only; each proposed mechanism lands as its own PR behind its own plan.

What it contains, beyond the brief:

• The original A–E catalogue validated by measurement (git 2.34.1 ✓; repos.yaml now 33; smoke_tests.txt 11 ✓) and enlarged out-of-sample with this campaign's own live errors (F1–F5): the shared-index sweep repeated ×3 in one day with a same-morning memory note about it, three piped-exit/silent-flag instrument errors, one wrong checkpoint claim published then caught by a deliberate falsifying probe, two stale-claim refusal catches, one API-gate false positive.
• The taxonomy compressed to two root causes: verification aimed at the believed path, and state consulted from memory instead of the world.
• The hypothesis tested and sharpened: refusals still catch ~100% of what they guard; informs still catch ~0% — including the agent's own live self-awareness (F3 was committed mid-campaign about that exact error shape). The dividing line is whether anything interrupts the action. Refusals now carry a measured false-positive cost column (F5).
• Six ranked mitigations with per-entry catch mapping, cost, and failure mode — led by set -o pipefail as the agent-shell default (deletes F2's class outright, one line) and a Mind commit guard (the E1/F1 four-incident class, converted from the memory note that failed).
• An honest verdict on memory (works consult-on-purpose, fails as tripwire — convert guard-shaped notes into refusals) and a first estimate of the unnoticed-error set (~7% caught-late rate among boundary-crossing claims → expect 1–3 undetected false statements live in the record).

Parallel PR from main (docs/ only; no file overlap with any active claim). Shipped under the human-authorized corrective-PR exception for Heart RED (standing this session). PR-open only; merge stays human.

API Changes

None — documentation only.

Test Plan

• PyAutoBrain suite in the worktree (gate leg).
• Every §1 validation number carries its command; §2 incidents cite the campaign issue trail.

Generated by the PyAutoLabs agent workflow.