G6 complete: Phase B measured (Phase A was survivorship bias) + the missing deployment docs

[SebastianKrn]

Closes milestone G6. Two independent tracks.

Track A — G6 Phase B

Amendment A2 fixed the method before any repair ran: mutually blind parallel repairers, repair may not add a test the model did not write, null output is ABANDONED at zero and must not be averaged in, truncated output is salvaged not completed.

	Phase A	Phase B
green cells	12/24	21/24
test methods	154 (3 red)	421 (0 red)
line / branch coverage	100 % / 100 %	90.5 % / 78.8 %
mutation score	99.2 %	73.2 %

The quality metrics got worse, and that is the finding. Phase A's perfect figures were computed only over the cells that happened to compile — exclusively the small controllers. The God class was never in the denominator. Repair puts it in. Phase A was survivorship bias, visible only because the protocol forced every failed cell to be measured to the end.

Repaired, the God class reaches ~83 % line coverage at a 44–56 % mutation score — half the injected faults survive the one class the experiment is about. A coverage gate would have passed it.

Also measured:

• one model wrote 134 test methods where the other wrote 13, at byte-identical value (21/21 lines, 13/13 mutants);
• zero BUG-FOUND across 15 wrong-expectation repairs — these tests pin behaviour, they do not audit it;
• the open-weight model emitted getStatusCodeValue(), present in corpus A's Spring 4.3 and removed in corpus B's Spring 7 — migrating to a very new stack makes LLM support temporarily worse.

Two defects in our own work, disclosed not fixed away:

• A3 — one cell in 24 was decided by our extraction rule, not the model: it wrote a draft, rejected it in plain text, then wrote a correct 26-test class; our "first fenced block" rule kept the draft. Not re-extracted; 12/24 stands. The failure taxonomy is corrected instead.
• A4 — the repair-effort clock is contaminated by A2.1's own parallelism (one cell: 24.7 wall-clock minutes for a build that worked 5.6 s). Minutes published as a measured upper bound; the transferable figures are the 52 fixes and their categories.

Adoption (ADR-0011): one class per unit, highest mutation score, ties to fewer test methods — the rule that rejects the 134-method class, and which selected a mixed set (2 frontier, 4 open-weight). 88 methods adopted → 99 module tests green, coverage 37.3 % → 81.3 %, ratchet 0.35 → 0.80. §3's 80 % target reached rather than declared for the first time; branch coverage still lags 23 points and the adopted God-class suite kills 44.1 % — both stated next to the number.

Track B — the deployment docs

ENGINEERING_STANDARDS.md §7 has always required docs/deployment.md. It did not exist — and the gap was not in DEVIATIONS.md either. Written now, every command executed before being written down, plus docs/MANUAL_TASKS.md as the by-hand checklist. Production deployment gets no invented steps: §10 states what does not exist.

Beginner-fatal fixes: --wait was missing from every human quickstart while all three CI workflows always used it; characterization silently ignored -Dtarget, so -Dtarget=modern went green against the legacy stand — now a fail-fast; modern verify needs a Docker daemon and nothing said so. Corrected: the legacy/ build failure is not the Java 8 source level — on JDK 26 compile and surefire succeed and it dies in maven-war-plugin:2.6 reflecting into java.util internals JDK 16 sealed.

Verification

char 47/47 vs legacy AND modern · e2e 34/34 vs legacy AND modern · harness 24/24 · both testbed validations · modern verify 99 tests at the 0.80 gate. The safety net is untouched.

🤖 Generated with Claude Code