Move three annotation blocks out of declaration bodies so main parses again

[gunbai-bot]

Why

main has been red since 523657c15c (#10630) — three consecutive witnesses runs, ~6 hours at time of writing. The cause is 17 parse refusals in dag/gunbc/machine_intake/megarac_media_attach.dag, all one class:

source annotation sits inside a declaration body. Only module-item grain is
modeled; move it above the declaration it describes.

That is DESIGN §4c working correctly, not a compiler defect — the initial .dag realization admits only standalone leading // blocks attached to module-scope declarations, and body-position forms refuse until separately modeled.

Why it matters beyond one file

The refusal cascades: the parse phase produces no index, so namespace-wave-admission reports NOT RUN and the floor lane fails two phases rather than one.

It also blocks pull requests, which read main as their base. #10921 fails with megarac_media_attach.dag does not parse at the base revision (17 diagnostic(s)), so its base-side declarations cannot be read — while its own measurement is clean (claims_failed=0 passed=3520).

What changed

Each block moved above its enclosing declaration, appended to the annotation already there and separated by a bare //:

block	moved above
212–218	func megarac_attach_remote_image
303–306	func megarac_attach_with_token
441–446	func megarac_start_and_confirm

The prose is preserved, not deleted. It is the irreducible rationale §4c exists to keep — one block records that malformed presented state is unknown rather than "not attached", naming the absorbing fallback §5 forbids. Deleting it to satisfy the parser would discard the reasoning and keep only the syntax.

No code is changed. Verified by counting both sides: 5 func declarations and 306 non-comment lines, identical before and after.

Verification

I could not verify the parse locally — the session container is arm64 and a remote ctrl-build dispatch syncs the repo to a committed SHA and runs git clean, so an uncommitted working tree never reaches the runner. My first attempt therefore parsed main, not this change, and proved nothing. This PR's own CI is the verification; the discriminating signal is the parse FAIL lines disappearing and namespace-wave-admission running instead of reporting NOT RUN.

Opened without waiting for an operator answer because a red main is a stopped line. I am not merging it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BJGbrvU2EgNeUiWfc5yK2c

[gunbai-bot]

Closing as a duplicate of #10930, which was opened five hours earlier and is smaller for the same result (+17/-17 vs +20/-17; I added // separators it did not need). Verified equivalent: both leave 0 body-position annotations, 5 func declarations, 306 non-comment lines.

The diagnosis is carried over to #10930 — the fix works (0 parse FAILs, claims_failed=0) and cannot green its own CI, because namespace-wave-admission reads the still-broken base. Deleting this branch with the close so no successor PR is minted.

— sent from merry-bear-25