docs: record measurement/instrumentation discipline in CLAUDE.md

[jkitchin]

What

Adds a Measurement & instrumentation discipline section to CLAUDE.md, plus targeted rules in Working on an issue, Workflow, and Key Constraints.

Documentation only — no code, no behavior change.

Why

Every rule is derived from a specific failure in recent sessions, not from general principle. Reviewing those sessions, the dominant failure was not a wrong measurement — it was a measurement that never happened and reported success anyway. One habit, many faces:

Incident	What it reported
Probe used Constraint.lhs/.expr; the real attribute is .body	traversed nothing, printed results
deepcopy(Model) raises TypeError (can't pickle PyModelRepr), caught by a bare except	a whole fallback was an invisible no-op, "working"
Attribution probe gated only the root dive	its "neither" arm was meaningless
Sparse patch landed in the algebraic extractor, not the repr one	"parity confirmed" was dense-vs-dense
Parity test never asserted issparse() on the forced-sparse arm	proved nothing, believed
pytest imported discopt from the main tree, not the worktree	19 bogus failures in a PR review
A background sweep declared dead while still running	two wasted runs, the second timing out on contention the first created

Two published claims were retracted on remeasurement: a "leftover model state slows relaxation ~40%" that was really +2% (pooled sd 4.55, with an unrelated pytest holding ~87% CPU), and a gastrans040 regression that was a timing artifact. Hence the load-gate / interleaved-control / report-a-spread rule, which explicitly covers stray load you created yourself.

Contents

New section (items 6–11, continuing the Development Philosophy numbering): prove the probe fired (executed-assertion counter, non-zero exit at zero); never swallow an exception in an instrument; verify which code you loaded (__file__ and a version-unique marker, asserted absent on baseline runs); timing claims need a load gate + interleaved control + spread; a long job needs incremental output and "no output" is not "dead"; retract your own published claims in writing.

Working on an issue gains item 6, finish rather than document — an investigation with three artifacts and no code change is the signal to start building.

Workflow gains three operational rules: no issue-closing keyword near a negation (GitHub matched close #863 inside "Does NOT close #863" and closed it); look up an API rather than guessing attribute names; re-derive order-of-magnitude figures before stating them (91.5 TB was written for 91.5 GB).

Key Constraints gains the two silent-corruption traps that have each already cost a wrong conclusion: np.asarray() on a scipy sparse matrix returns a 0-d object array without raising, and INF in the Rust LP layer is the sentinel 1e20, not f64::INFINITY — so unboundedness must be tested on the bound, never on a product. That second one is not hypothetical: it is the root cause of the false optimality certificate fixed in the companion PR.

Verification

pre-commit hooks ran on commit (ruff / ruff-format / mypy / cargo fmt — all skipped or passed; no code files touched).