S6: validation report + figures + MkDocs Validation page + parity README; remove legacy zdt/ (milestone-12a)

[CooperBigFoot]

s6-reporting (milestone-12a) — validation report, docs, legacy removal (FINAL step)

Turns the committed 30-seed results JSON into the human-facing report + a citable docs page, and removes the superseded legacy suite. Completes the milestone.

Files

• benchmarks/render.py (NEW): JSON → markdown tables (injected between <!-- BEGIN:id -->/<!-- END:id --> markers) + matplotlib figures. Carries a fail-fast consistency gate (check_report_consistency) called inside render() before any file write — it RAISES if the report prose's claimed-indistinguishable/at-least-as-good set contradicts any committed verdict cell, so the citation page can never self-contradict.
• benchmarks/reproduce.py (NEW): the one uv run python benchmarks/reproduce.py regenerate-everything entry (sweep → render).
• benchmarks/README.md (REWRITE): canonical report, tables auto-rendered from the committed JSON + the 3 figures embedded.
• README.md (root): the ## Benchmarks link replaced with one honest parity paragraph + links.
• mkdocs.yml + docs/validation.md (NEW) + docs/assets/validation/*.png: a citable Validation page (passes mkdocs build --strict).
• Removed the entire legacy benchmarks/zdt/ tree (problems, run_benchmark.py, old results/png/log) — superseded; nothing imports it.

Report framing (parity, not superiority; reconciled to the committed 30-seed verdicts)

• SO: indistinguishable from pymoo and DEAP on all six functions (objective_error + solution_distance). success_rate documented as a strict-threshold non-convergence property at this budget (0 for all three — the error metrics carry the SO evidence).
• MO: indistinguishable on ZDT1, ZDT2, ZDT3, DTLZ2 (IGD+, GD, hypervolume). ZDT4/ZDT6 documented as the two in-our-favour exceptions: none of the three converges at this budget and ctrl-freak's IGD+/GD/HV is at least as good as both baselines (ZDT4 HV degenerate 0/0/0).
• The old "pymoo ZDT3 scatter" caption is removed and replaced with the real cause: the legacy baseline applied PM per-individual (~0.03 vs ~1.0 expected gene flips, ~30× under-mutation) — fixed in s4b.

Acceptance (verified twice — critic + orchestrator)

• render.py (gate passes, 7 tables + 3 figures), mkdocs build --strict, ruff, ty check src/ — all pass.
• full uv run pytest: 525 passed @ 98.89% (the s5 doctest gate now also covers render/reproduce). Legacy zdt/ gone; nothing imports it.

Plan converged via planner↔critic over 2 rounds (round 2 added the prose↔verdict consistency gate; critic verified it passes on the committed JSON and raises on a flipped cell).

[CooperBigFoot]

Adversarial review: APPROVE (final step, s6-reporting)

Audited the committed diff and verified everything by execution in the worktree.

Consistency gate (the citation shield)

• Passes on the committed 30-seed JSON; gate is fail-fast — render.py:622-631 runs build_tables → check_report_consistency → inject_tables → figures, before any file write.
• Raise-on-tamper proven: a scratch JSON with zdt1/igd_plus flipped to "ctrl-freak worse" makes render() raise ValueError and leaves README.md + docs/validation.md byte-unchanged (sha256 before==after).

Tables trace to JSON (no hardcoded numbers) — sphere/objective_error, zdt4/igd_plus, zdt4/hypervolume (degenerate), zdt6/hypervolume, schwefel/solution_distance, zdt2/zdt3 igd_plus all match committed verdicts.

Prose ↔ verdicts — SO 6/6 equivalent on both error metrics; MO zdt1/2/3/dtlz2 equivalent on all three metrics; ZDT4/ZDT6 framed only as in-our-favour exceptions (grep "worse" → none). No superiority framing. ZDT3 'visible scatter' caption replaced with the real cause (legacy per-individual PM, ~30× under-mutation).

Legacy removal — benchmarks/zdt/ tree gone; no real imports remain. Do-not-touch set (problems/metrics/harness/run.py/stats.py/config/tests/pyproject/uv.lock/.github) untouched.

Acceptance — render gate PASS + deterministic (figures byte-identical on re-render); ruff/ty(src) PASS; doctest gate 11 passed (collects render+reproduce); full suite 525 passed @ 98.89%; mkdocs build --strict exit 0.

CI: 4/4 green (py3.11/3.12/3.13 + numpy-floor). mergeStateStatus CLEAN.

Merging — completes milestone-12a (8/8 steps).