Close M17, and make EVAL.md readable in order

[jadoon200]

Three loose ends: a stale private plan, an untested milestone, and a document grown past the
point where it could be trusted to agree with itself.

M17 / Q3 — a foundation model changes nothing that matters

Chronos-Bolt zero-shot, Apache-2.0 per the licence audit, kept as an optional extra —
torch has no place in a container that serves a precomputed snapshot.

The ceiling is the finding. Chronos-Bolt is trained on quantile levels 0.1–0.9 and
cannot express p95 or p99. It clamps to p90 and warns rather than failing, so a newsvendor
sizer asking for p95 receives p90 wearing a p95 label — worse than an error, because it looks
like an answer and nothing downstream can detect it. For a project whose thesis is that a
controller consumes a calibrated quantile, the most capable forecaster available cannot
supply the object the decision layer needs. Pinned by test.

Accuracy — marginal, exactly as registered. Wins 1 MASE row of 5; loses to plain
seasonal-naive on 3.

The decision — identical everywhere. Not one cell changes; it reproduces the classical
result on every workload including the materna-2 exception.

workload	classical best	chronos-bolt
materna-2	2/4	2
the other four	0/4	0

Five forecaster families have now been tested against the same decision and none moves it.
That retires the caveat left on 08-08 that "most sophisticated in this repository" wasn't
the field's best. The commitment result is a property of the demand, not of the predictor.

EVAL.md, read as a hostile reviewer

Three genuine incoherences, all from sections written days apart:

1. The opening paragraph still said the newsvendor metric would arrive "once M4–M6 exist"
   — a week after they did.
2. A line in the 08-07 section still said Q13 remained open, which two later sections
   refute.
3. The M4 gate section presented a green verdict with no hint that this exact gate passed
   for five days while missing the actuation bug. That's the single most important
   cross-reference in the document and it was absent.

All three now carry forward pointers, and the document opens with a map: it is chronological,
later sections overturn earlier ones, superseded tables are banner-marked, and every question
is listed with where it ended up. Internal anchors verified to resolve.

Engineering note

Installing the extra made the suite segfault — torch and lightgbm each load an OpenMP
runtime and on macOS the pair crashes the interpreter mid-run (201 pass without, 3 pass alone,
both together die). Pinned to one thread rather than reaching for KMP_DUPLICATE_LIB_OK,
which suppresses the duplicate-runtime check and can corrupt memory silently. Foundation tests
are marked and excluded from the default run; CI never installs the extra and skips cleanly.

Result

Served findings 9 → 10. .claude/PLAN.md rewritten to match reality (gitignored, so not
in this diff).

Gate

ruff, format, mypy, pytest, validate-simulator, make evaluate-foundation — all verified by
exit code.