Check curve, ablation and baseline completion in code

[drewOrc]

Why

A shell loop once passed make a single argument, curve MODEL=bert. Make read it as a variable assignment, ran its default target (setup) and exited 0, so two curves were reported done with no point run. No result was affected, but exit 0 is not evidence that anything ran. This makes each index command prove its own completion.

What changes

New src/tinyrouter/completeness.py, used by curves.run_curve and baselines.run_all.

Command	Before running (plan)	After running (index read back from disk)	Printed only when both pass
make curve MODEL=bert|modernbert	exactly 6 k x 3 seeds of that model, OOS default, each once	same 18 (k, seed), each once; every archive's SHA-256 equal on disk, in the manifest and in the index	completed 18/18 encoder points (bert)
make oos-ablation	ModernBERT, k=100, 0 OOS rows x 3 seeds, each once	same 3 points, same SHA-256 check	completed 3/3 ablation points
make baselines (and the first half of make curve)	2 baselines x 6 k x 3 seeds, each once	same 36 (baseline, k, seed), same SHA-256 check	completed 36/36 baseline points

• The plan check runs before any training, fitting or data loading.
• The index is written to <name>.json.tmp, read back and verified, and only then moved into place; on failure the temp file is removed, the previous index stays as it was, and the command raises (non-zero exit).
• The expected sets are literals in completeness.py, not read from CURVE_KS / SEEDS, so a wrong loop cannot agree with its own check. A test pins that the two agree today.
• make baselines no longer prints wrote N baseline runs; the completion line replaces it.

No change to docs/PLAN.md, the acceptance criteria, hyperparameters or any result file. No training or paid API call was run.

Tests

27 new tests in tests/test_completeness.py: for the curve, the ablation and the baselines, a missing point and a duplicated point at the plan check and at the index check (12), plus the completion message and its ordering, archive changed / manifest disagrees / archive deleted, wrong model, OOS rows in the ablation, and a failed rerun leaving the old index intact. Three existing tests in tests/test_curves.py ran a 3-point subset of a curve; they now run the full curve and filter to k=100.

Mutation check (first commit): each of 13 mutations (dropping each plan check, each index check, the manifest comparison, the missing-file check, duplicate / missing / unexpected detection, verifying after the move instead of before, printing the message before publishing) turned at least one test red; all restored.

Review follow-up (second commit)

• No default target. The entry point of the incident still exited 0: make -n "curve MODEL=bert" expanded setup. The Makefile now stops with $(error ...) when MAKECMDGOALS is empty, before any target. Chosen over .DEFAULT_GOAL := help with a failing help target because it fails at parse time for every no-goal call, including make -n, and its message names the actual mistake. README and CI only ever call named targets. tests/test_makefile.py (dry runs only, inherited MAKEFLAGS dropped): the quoted form, a bare make and make MODEL=bert exit non-zero without reaching setup; make -n curve MODEL=bert and make -n setup still exit 0. Make versions differ on the quoted form: GNU make 3.81 (macOS, where the incident happened) reads it as an assignment and now hits the guard; make 4.x (CI) reads it as an unknown target and fails with No rule to make target. The quoted-form test accepts either refusal, and the guard itself is pinned by the bare and assignment-only calls, which every version treats the same.
• Read-back is pinned. A test makes the write of the temporary index lose a point while the body in memory stays whole; the check must fail.
• Shared archives. check_archives refuses two points listing the same archive file or SHA-256, so the baseline index gets the check the curve index already had.
• Docs. Makefile comments and README: decide completion by matching the whole line completed 18/18 encoder points (<model>), since make curve prints the baselines' completed 36/36 baseline points before the curve runs.

Mutation check for these: removing the guard, making it fire on every call, verifying the in-memory body instead of the file, and dropping the shared-archive check each turned at least one test red; all restored.

make lint, make test (348 passed, 1 skipped, 7 network deselected) and make smoke pass locally.