Measure Go's elidable material, and correct the sequencing warning

[ojassug]

Precondition check for widening elision past TypeScript/JavaScript/Python — the one roadmap item still open. Measured before writing any of it. Docs only; no source changes.

The material

Ceiling = share of bytes inside func bodies clearing the shipped filters (MIN_REGION_BYTES 104, isSubstantiveRegion) on scanBraceSpans's between-brace boundary. TypeScript and Python measured with the shipped selectElisionRegions over the frozen corpus.

corpus	files	ceiling, non-test	ceiling, all	median/file	no region
Go — app (cli/cli, cobra, gin)	1,028	65.36%	81.44%	63.3%	9.5%
Go — stdlib (golang/go src/)	5,387	54.78%	59.81%	39.5%	28.2%
TypeScript — this repo	62	57.78%	—	58.6%	11.3%
Python — pip	45	46.88%	—	53.8%	2.2%

TypeScript turns 57.78% into 24.56% achieved at target 0.3, so Go projects to 23–28% — at or above the product's best language. The precondition holds.

The sequencing warning was wrong, in the dangerous direction

All three documents said "add the scanner alone and §33's measurement gate refuses the item, converting a 0% into a fallback." Measured, that holds only for a file with no struct, class or import:

case	symbolsBefore	S_k	astMeasured	measurementGate	fallback
Go with struct/import	type:Point, import:fmt	0.0000	true	pass	false
Go without either	(none)	0.0000	false	refuse	true

extractSymbols yields no function symbols at all for Go, C, Java or Rust — type:Point and import:fmt are incidental matches by the TypeScript regexes, and both survive body elision by construction because signatures are retained. On real source, which nearly always carries a struct or import, scanner-first passes every gate while measuring nothing. Every function body could be deleted and S_k stays 0.0000.

This is C1's shape one step over, and §33 does not cover it: §33 closed "the before-set is empty so R_AST defaults to 1.0"; this is the sibling where the before-set is non-empty but structurally incapable of registering the loss.

Order, for that reason: extractSymbols → validator → REGION_ELISION_LANGUAGES + scanner. Step 1 alone is a free negative control.

The claim is struck through in all three documents rather than deleted, with the measurement beside it — including text added earlier the same day. Leaving a disproven statement standing in a third document is this project's own recurring failure.

Two findings worth more than the headline

Test files are the larger prize. _test.go is 53 MB against 36 MB of source in the app corpus, at 92.22% elidable with 0.7% having no region. Nothing in this project has been counting them.

The ceiling is not the constraint; the gates are. Target 0.9 gives TypeScript 21.37% with 25 files unchanged, against 24.56% with 12 at target 0.3 — pushing harder trips the constraint and drift gates. §48 reproduced; it bounds Go too.

Method

The cross-check moved the number and that is why it was run: app-only read 65.36%, the stdlib pulled it to 54.78%, and the cause was checked rather than averaged — 21.7% of stdlib source bytes are in files with no elidable region, dominated by generated tables the DO NOT EDIT filter missed (opGen.go alone is 3.99 MB). One corpus would have overstated this by ten points.

The instrument was validated before the result was believed: 12/12 cases including raw-string literals, both comment forms, interface methods with no body, and closures counted once.

Not established: the 23–28% projection borrows TypeScript's conversion factor, which embeds TypeScript's fallback rate. Go's own is unmeasurable until the validator and extractSymbols exist.

669 tests, typecheck, lint and build green.

🤖 Generated with Claude Code