test: Falsifiability A — skip-list ablation experiment (mache-3509aa)

[jamestexas]

Summary

ADR-0013 named two first-class falsifiability experiments. This is A — the headline.

Hypothesis: with `_lsp_refs` populated, dropping the `dead_code` skip-list flags zero NEW dead constructs that LSP indexed. Any new finding must be a method on a type LSP didn't see (genuine coverage gap — wedge case 1 from the ADR).

Procedure

1. Run `dead_code` in production form (skip-list intact). Findings A.
2. Run `dead_code` with the `skipped` CTE neutralized. Findings B.
3. Diff = B \ A — newly flagged dead.
4. For each row in diff, look up its `node_id` in `_lsp_refs`. If LSP HAS a reference to that def, the design is broken — the alive-check's binding arm should have made it alive.

Implementation: stripSkippedCTE

Walks the rule body with paren-depth balancing to locate the matching close paren of `skipped AS (...)` — the inline SQL has nested parens for the `IN ('main','init',...)` list, so naive regex matching fails. Replaces the whole CTE with an empty form: `skipped AS (SELECT '' AS node_id WHERE 0)`.

Tests

• `TestStripSkippedCTE_RealRuleBody` — introspects the production registry, transforms the rule, asserts the skip-list (`'ServeHTTP'` marker) is gone AND both queries are still valid SQL. Future refactors that move the skip-list out of an inline CTE fail this test loudly.
• `TestFalsifiabilityA_SyntheticHarness` — fixture with two methods on `MyType`: `Read` (LSP indexed) and `Quux` (LSP didn't). `Read` is in the skip-list and has an LSP binding row → not flagged in A; ablation drops the skip-list, alive-check's binding arm matches `target_node_id` → not flagged in B either. Diff doesn't contain `Read` — the headline. `Quux` is dead in both paths.
• `TestFalsifiabilityA_IntegrationOnLeylineParse` — gated on `MACHE_FALSIFIABILITY_INTEGRATION=1` + `leyline` binary. Runs both variants on a real LLO-built mache `.db`, then for each newly-flagged def looks up `_lsp_refs` by token. If LSP HAS a reference but the rule flagged dead, the test fails with the suspect list.

Note

Tests pin `db.SetMaxOpenConns(1)` so TEMP views from `ensureCanonicalViews` remain visible to subsequent queries. Without the pin, `*sql.DB` pools fresh connections and the view-creation isn't seen by later `db.Query` calls.

ADR-0013 closeout

This is the last bead. All 5 migration steps + 2 falsifiability experiments now shipped:

Step	Bead	PR
1 — producer-side token extraction	`ley-line-453f7e`	LLO @`12330fc`
2 — `_index_coverage` table	`mache-346cfe`	#340
3 — canonical views (mention)	`mache-346d0f`	#341
3 — canonical views (binding)	`mache-346d0f`	#347
4 — rule rewrites	`mache-346d1e`	#346
5 — drop LIKE fallback	`mache-346d2b`	#348
Falsifiability B — π round-trip	`mache-354464`	#349
Falsifiability A — skip-list ablation	`mache-3509aa`	this

Closes `mache-3509aa`.

[github-actions]

find_smells (advisory)

Scoped to files changed in this PR. Rules below run on the standalone (no-LLO) cross-ref tables; _ast rules (cyclomatic_complexity, long_function, long_file, magic_int_in_comparison) are not exercised here.

fan_out_skew — 2 finding(s) in changed files

Source	Node	Metric
cmd/serve_lsp.go	cmd/functions/enrichAndQueryDiagnostics/source	19
cmd/serve_lsp.go	cmd/functions/enrichAndQueryTypeInfo/source	19

Rules: see docs/ARCHITECTURE.md for the full registry. Advisory only — these are heuristics, not gates.