#135)

Follow-up to #134 — the maintainer pair agreed offline that the 4-row +
Demos compromise still leaves audit-flagged claims visible. This PR
removes the Demos section entirely and trims Receipts to the three rows
that share one theme: each is a byte-precise CI gate that came out of an
audit-named gap.

What changes vs #134:
- Receipts becomes 3 rows: sensor parity (#125), audio parity (#121),
  training-data lock (#130). All three are CI-gated, all three respond
  directly to a specific finding in the 2026-05-03 staff audit.
- Demos section deleted. The hackathon-phase claims (BCE 0.7698, 5/5 spot
  checks, sub-2 ms sensor latency) are NOT demoted; they are removed.
  Auditing them as "illustrative" still left the audit-flagged numbers
  on the front page; deletion is honest.
- New "Engineering" subsection holds the diagnostic-but-not-strict-receipt
  facts: workspace gates green from clean checkout (#127), codec
  independence enforced by lint:deps (#126), self-contained loupe HTML
  ~ 117 KB.

Why: per the audit memo + the maintainer-pair review on PR #103, four of
the original six receipts were misleading or unsupported. The right shape
is "every Receipts row is a byte-precise CI gate" — anything diagnostic
goes elsewhere or doesn't appear at all. Loupe HTML's 117 KB is a real
reproducible measurement but not a CI-gated parity check, so it lives
under Engineering rather than Receipts.

Validated:
- pnpm exec prettier --check README.md → green
- git diff --check → clean
- All four PR cross-links resolve to merged PRs (#121, #125, #126, #130).

Per ADR-0013: README.md is doctrine-bearing per §1; this is drift
correction tightening the gap between what's claimed and what CI gates
prove. Cannot self-merge — needs an independent review pass.