[GOLD-RING] compiler: compound assignment and hoisted nested fn

.claude/skills/oracle-method.md

@@ -0,0 +1,90 @@
+# The oracle method — reconstructing a published table from its record
+
+Distilled from the TNF audit (W846–W876): twenty tables put under regenerators,
+sixteen document defects found and fixed, five of my own claims withdrawn along
+the way. Each rule below was paid for by a concrete failure named beside it.
+
+## When to use
+
+A document prints numeric tables; machine-written records exist (or should).
+Before trusting a number, before asserting a defect, and before shipping a new
+record of your own.
+
+## The order of operations — this exact order
+
+1. **`t27c known --dir <docdir> --about <X>`** before measuring anything about X.
+   Gates, baselines and captions answer most questions in a minute. Paid for by
+   W870: 55 place-and-route runs rediscovered a defect whose figure sat in one
+   baseline line. Signal weights: baseline = strong, gate = medium, caption =
+   weak (a caption *naming* a table is not prior art about your claim).
+
+2. **Read the record's own metadata** — `description`, `seed`, `n`, `method`,
+   generator name. It outranks every similarity heuristic. Paid for by
+   `breakeven`: every score pointed at the wrong table; the JSON's own
+   description field named the right one.
+
+3. **Never map by similarity.** Substring stems voted wrong on a third of one
+   corpus; numeric overlap over-reported 86 % where reading gave 8 %; a perfect
+   16/16 mapping once sat at F=0.155 because the record stores 16 digits and the
+   table prints 3. Similarity measures formatting overlap, in both error
+   directions. Reconstruction is the only verdict.
+
+## Writing the regenerator
+
+- **The table is a view.** Either every printed cell comes back at *printed*
+  precision or it does not. `0.282` asserts three decimals; a 1 % band passes
+  values the table does not claim.
+- **Match rows by key, never position.** A selection makes position meaningless
+  (33 phantom mismatches from zipping σ=1.5 against σ=1.0).
+- **Identify columns; never assume them.** Assert an identity —
+  `ratio == a/b` to 1e-9 — before judging any cell (38 phantom mismatches from
+  assuming column 3).
+- **Guard the vacuous pass.** Pin the expected printed-row count; an emptied
+  table must fail, not pass 0/0. Found only by mutation testing.
+- **Caption claims are data.** Recompute counts, sample sizes, superlatives —
+  *conditionally on the phrase being present in the source*, so a later fix
+  turns the light green. A check that cannot pass is not a check.
+- **Report the selection.** If the record holds more rows than the table prints,
+  list the unprinted rows with one salient value each. Two tables hid their most
+  damaging rows this way (a 2.48e+35 blowup one step past the printed sweep; a
+  competitor's 4×-better cell).
+- **Formatting is data.** Bold/dagger marks encode a criterion; recompute it
+  (12/4/14 at the record's own tolerance caught a parser that dropped rows).
+- **Do not edit the document from the oracle.** Report, exit 1, fix separately.
+
+## Verifying the regenerator — adversarially, always
+
+Run it, read it for the cheats above, then **mutation-test on scratch copies**:
+perturb one record value → must fail; delete a printed row → must fail; empty
+the table → must fail; restore a fixed defect phrase → must fail. Two real holes
+in five scripts survived ordinary review and fell to mutations in minutes.
+
+## Reporting
+
+- Verdicts agreeing across instruments that share a blind spot are one verdict.
+  Diversity of instrument beats count of readers.
+- A reconstruction proves the table matches its record, **not that the record is
+  right** — the reach column matched its record perfectly and both were wrong;
+  only an independent definition (the paper's own proposition + the oracle)
+  broke the tie. Two signals derived from one quantity are one signal.
+- Report what *survives* beside what falls (a neighbouring claim recomputed true
+  while its sibling fell — that distinction is what separates an audit from a
+  hit job), and score your own record: which claims are about the document,
+  which about your method; the withdrawal count is the number to watch.
+- A readiness percentage that averages progress cannot see a binary gate.
+  Report the conjunction of the gates.
+
+## Shipping your own records
+
+Ship the generator in the same commit; state which fields reproduce on a re-run
+and which pin the shipped file (nextpnr logs hash differently every run — a
+non-reproducible hash is a trap for the next reader). A record no script can
+rebuild cannot be corrected at source.
+
+## Tools
+
+`t27c known` (prior art), `t27c provenance` (three weighted signals, UNDECIDED
+on divergence, per-column mode for multi-record tables), `t27c battery` (run
+every oracle and gate with per-script exit codes — never read `rc=$?` after a
+pipeline), `t27c editcheck`, `t27c recompute-diff --label` (whole-file mode is a
+false-pass generator).