Registry: the lock closes every form and resolution is proved by a corpus

[the-homeless-god]

Two P1 tasks about the registry: 2031 (no form of a dependency reaches the lock) and the remainder of 4909 (properties 2 and 3 of ADR-0059 section 7). Point 4 of 4909 was closed earlier today as false from birth and is not touched here.

4909, properties 2 and 3

A new guard scripts/guards/registry-resolution-guard.fscript plays a corpus in packaging/registry/cases (a ledger of seven rows and six cases written as key = value lines).

• Property 2, the answer does not depend on the order of the ledger rows. Every case of kind "choice" runs on all rotations of the ledger plus the reversed order — seven rows give eight orders — and the answers are compared byte for byte. Run: cases 6, ledger orders 20, removed chosen entries 4, exit 0.
• The mandatory negative control is inside the check, not beside it. One answer over eight orders also goes green on an instrument that never reads the ledger, which is the mistake the measurement of the ADR fell into once. So for every choice case the guard drops one chosen entry in turn and demands the answer change; all four removals change it. Measured by hand as well: cutting Логика | 1.2.0 out of the corpus ledger turns the guard red, exit 1, two findings, the answer becoming Логика 1.0.0.
• Property 3, a refusal is named rather than swallowed. The four requests of the ADR table are corpus cases of kind "refusal", and both the code and the reason text are compared character for character. The run gives the promised FLANG_REESTR_KONFLIKT, FLANG_REESTR_KONFLIKT, FLANG_REESTR_ZAPROS, FLANG_REESTR_STROKA and four texts that agree with section 5.
• Forgery. The Forgery plan strips the list of ranges and versions out of every expected reason and appends a word to every expected answer: exit 1, six findings, every case named.
• Property 1 was repaired by the ledger reprint earlier today; here it gets its own shortcut registry-prints:check and its own called probe, so the fingerprints are checked by a job and not by hand. Negative control: one digit bent gives "not matched: 1, matched 6", exit 1.

The CI job registry-resolution runs the probes before the checks. Both checks also join the pre-push hook (4.5 s and 3.4 s here).

2031, the lock

Two edits, both in flang/src/emit/c/flang_repl.c:

1. lock_scan dropped an import without a "from" key by a bare continue. It is now resolved by repl_find_module — the same places and the same order the loader uses, because a list of places of its own would mean the lock records a file other than the one that will build.
2. lock_collect now opens a .flang-package whole through pkg_take_nested, the same body a package over a package uses, instead of handing it to the flang parser.

Runs after the fix, binary 0.7.24:

form	before	after
by name, orders-api.flang	exit 0, 149 B, 0 modules at 4 files	exit 0, 109766 B, 3 modules at 4 files
package by path, shop.flang	exit 1, FLANG_PARSE	exit 0, 1 module at 2 files
path import, hmac.flang	exit 0, 87465 B, 1 module at 4 files	exit 0, 168237 B, 3 modules at 4 files

The third row is the control and it moved too: sha256.flang inside the closure of hmac.flang imports Numbers and UTF-8 by name, and both were lost silently. So "one module" in the old measurement was not the boundary of the path form but the same defect inside it.

The end to end run of the task is green: lock shop.flang into an empty directory, flang check there gives exit 0 and "two files with the imports", and the same directory without the lock refuses with FLANG_PACKAGE, exit 1.

The instrument is scripts/guards/lock-closes-every-form.fscript: over all three forms it compares the length of the модули list with the file count check names, minus the entry file. Run: forms 3, exit 0, 9 s (check --fast names the same count but takes 3.5 s instead of 149 s on hmac.flang).

Two probes, not one. The cheap one replaces the measured length by zero. The honest one restores the old behaviour in the seed (wanted_bytes != 0 → == 0), rebuilds the binary and demands that lock-forms:check itself go red: measured here as "by name 0 modules at 4 files" and "path 1 at 4", exit 1, while the package form stays green because its repair is a different one.

Delivery

The edit fell only into flang/src/emit/c, so the seed is refreshed the fast way in its own commit, not reprinted. flang/self is untouched, so this needs no print batch.

The fingerprint names the commit of this branch, so it dies on the next rebase: the trunk then needs one line of repair, bootstrap-reprint.sh --perekommit. Same note as the reseed of the LSP fix earlier today.

Counted marks follow by instrument, not by arithmetic: three line marks in docs/tree-inventory.md (C lines +100, bootstrap +50, emit runtimes +50) and one row in scripts/ledgers/proved-share-ledger.txt. DESCRIPTION.md lost the limitation "a module found by name is not written to the lock" because it is no longer true.

All 24 pre-push checks green, 61 s.

🤖 Generated with Claude Code

https://claude.ai/code/session_01X5n29umGpiervMVgnvvqht