docs(efficacy): round 4's report and verdicts (#1875) (#1876) (#1949)

[braboj]

Closes #1875, closes #1876, closes #1949. Replaces #1953.

What

• docs/audits/2026-10-17-efficacy-thin.md and docs/audits/2026-10-17-efficacy-weaker.md: report.py output for each round-4 root, unedited. report.py names both 2026-10-17-efficacy.md, so each is renamed with a run suffix.
• docs/audits/2026-10-17-round-4.md: the verdicts on design §12.3 and §12.4. It has a growth table for every registered row, and thin hybrid's registered scores beside the key-supplied re-run (Settle whether the application must start without a secret key #1949).

Verdicts

Hypothesis	Verdict	Rows that count
12.3, thin spec (#1876)	Supported	hybrid, judged security +1.45; full +27 pp and hand +32 pp, security probes
12.4, weaker model (#1875)	Supported	full, judged security +0.56; hybrid, judged data protection +1.0; hand +18 pp, security probes

• Task success didn't grow under haiku. Full, short and hybrid scored below none.
• Both reports owe an escalation to K = 5, which hasn't been run.
• Under the thin spec, hybrid's 42% build suite and its probe rows measure the missing TARIFF_SECRET_KEY, not its code. With the key supplied, its suite scores are 87.1%, 94.6% and 84.8%.

How it was produced

• Both roots were judged three times per trial by claude-opus-5-5, with no cross-check (§12.2). Thin hybrid-2 was judged after fix(efficacy): keep a virtualenv that lost its pyvenv.cfg out of the judge's bundle (#1950) #1951.
• The archives were verified from a fresh clone of braboj/tariff-runs (rounds/2026-10-16-thin, rounds/2026-10-15-weaker, rounds/2026-10-16-thin-sensitivity).
• report.py ran on main at 8f09318.
• Growth is round 4's mean paired difference against none minus round 3's, as 2026-09-27-efficacy.md prints it. A row counts where that reaches the registered bar and round 4's verdict is "better".

No CHANGELOG entry: these are benchmark reports, not a template change. run_smoke.py 33/33; run_conformance.py 0 failed.

🤖 Generated with Claude Code