Record post-merge external findings report; fit the brief to two pages

[batbrainy]

Closes #41

What this is

The final verification record. The checklist preamble and the container-kill erratum both deferred to an "external findings report" that did not exist in the repository — this PR adds it, recorded from an actual authoritative run, and closes the one remaining render gate.

Changes

docs/evidence/2026-08-01-post-merge-verification.md — the external findings report. One uninterrupted scripted pass (2026-08-01T14:22:35Z → 14:30:35Z, 8m00s) of checklist §1 against 88e2260c from a fresh clone into a temporary directory:

• Both cold build paths (build --no-cache --pull, then image rm + up --build); exact service topology; readiness.
• Live reach without a token: one real poll — 100 events received, 97 push events created, budget debited.
• Empty-volume fixture phase: 4 created / 3 quarantined / 1 ignored, complete 2 / permanent_failure 1 per class, SQL state exactly 4/3/3/3/3, replay absorbing 4 duplicates with no reactivation.
• Three consecutive green suite runs (including fixed seed 4242): 1,734 + 10 examples, 0 failures each, with development-database observations unchanged (§16 test isolation).
• script/verify_recovery.sh --confirm full run: 45 checks passed, 0 failed — the re-run against the final SHA that the container-kill erratum demanded.
• Restart and down/up persistence; §4.5 budget-isolation probe (budget row byte-identical across 9 endpoint calls); §6 brakeman (0 warnings) and bundler-audit (clean) in the container; full-history secret scans, stale-doc scan, and a hand-reviewed forbidden-claim scan (24 hits, all negations).

Reference fixes — the checklist preamble and the erratum now link the recorded report instead of naming an external document that could not be read.

docs/DESIGN_BRIEF.md — tightened by ~65 words (955 → 891) so the §5 render gate passes: at US Letter, 11pt, 0.75-inch margins the PDF is exactly 2 pages with the final paragraph on page 2. All five required topics remain under their original headings; the tradeoffs table became an equivalent prose ledger. No claims were strengthened anywhere — the honesty language is intact and the §7 scan over the changed files still returns only negations.

Tradeoffs

• The report attests runtime behavior at 88e2260c; this PR is documentation-only on top of that SHA, which the report itself states — every command a reviewer runs executes the verified runtime unchanged.
• The QA PDF is not committed (per the checklist); the report records the exact render parameters so the two-page result is reproducible.

Verification

• The report is the verification: a failing assertion would have aborted the scripted run at that gate.
• CI runs on this PR; changes are Markdown-only.