reanalyze: duplicated batch/reactive pipelines have silently drifted; the editor runs the untested one

DeadType.ml、ReactiveTypeDeps.ml、そして reanalyze_server.ml:309 のパスから始めて、再エクスポート依存関係がどのように扱われているかを比較します。make -C tests/analysis_tests test を実行し、その後、デフォルトではない reactive target と ci.yml:201 を調べます。チェックイン済みの corpus に、TypeReexport の fixture を含む、batch と reactive の一致する finding set が含まれ、CI がその一致を確認すれば完了です。

reanalyze computes dead-code liveness twice: a batch pipeline (Dead_type, Liveness, the frozen stores) and a reactive one (Reactive_type_deps, Reactive_liveness, ...). They are meant to agree. They do not, and the way they drifted is worth recording, because the same thing nearly happened again last week.

Which path do users get?

Not the one CI tests.

(* reanalyze_server.ml:309 *)
(* Editor mode only: the server always behaves like `reanalyze -json`. *)
Cli.reactive := true;   (* Force reactive mode in server. *)

Cli.reactive defaults to false (cli.ml:29), so:

who	mode
editor, via rescript-tools reanalyze-server	reactive
rescript-tools reanalyze on the CLI	batch
CI	batch

ci.yml:201 is the only place the suite runs: make -C tests/analysis_tests test, which reaches tests-reanalyze/deadcode/test.sh — batch, no -reactive. The two targets that would touch reactive, test-reactive-server and test-reanalyze-order-independence, are not default targets and appear in no workflow; their Makefile says "not run by default (takes longer)". So interactive users run the untested path.

The bug

On master, on the checked-in corpus:

batch     405 issues
reactive  412 issues

All 7 extra reactive findings are false positives on type re-export equations (type y = x = {...}):

TypeReexport.res:27:5-21   UseReexported.originalType.usedField is a record label never used to read a value
TypeReexport.res:48:5-24   UseOriginal.originalType.directlyUsed is a record label never used to read a value
TypeReexport.res:70:5-21   OnlyReexportedDead.originalType.usedField is a record label never used to read a value
TypeReexport.res:91:23     VariantUseReexported.originalType.A is a variant case which is never constructed
TypeReexport.res:109:23    VariantUseOriginal.originalType.A is a variant case which is never constructed
TypeReexportCrossFileA.res:5:3-19  originalRecord.usedField is a record label never used to read a value
TypeReexportCrossFileA.res:0:1     TypeReexportCrossFileA is a dead module as all its items are dead.

Every one of those labels is read, through the re-exporting type. The fixture states the expected behaviour in its own comments:

// Expected: originalType.usedField = LIVE (propagated)
//           originalType.unusedField = DEAD
//   warnings: only originalType.unusedField should be reported

Reactive says the opposite, and reports a whole module dead. This is what the editor shows.

Cause, confirmed by A/B

Disable only the manifest/re-export block at the end of Dead_type.process_type_label_dependencies and re-run the batch pass:

batch without manifest linking   412 issues
does batch-without-manifest == reactive ?  True
  reactive-only vs no-manifest: []
  no-manifest-only vs reactive: []

Set-identical in both directions. The reactive pipeline is missing precisely that linking and nothing else.

How it got there

date
2025-12-28	#8092	reactive pipeline lands, incl. ReactiveTypeDeps.ml
2026-01-09	#8127	reanalyze-server lands; the editor starts using reactive
2026-01-23	#8217	re-export linking added to DeadType.ml only

ReactiveTypeDeps.ml already existed when #8217 was written — it is in the tree at that commit, with twelve sibling Reactive* files. #8217 touched DeadType.ml, DeadValue.ml, DeadCommon.ml, Decl.ml, three new fixtures and expected/deadcode.txt, and no reactive file. The feature was one-sided from birth and has been for ~7 months.

Nothing caught it because the same commit added the fixture and regenerated expected/deadcode.txt in batch mode. The fixture that exists to pin this behaviour passes, in the mode nobody ships.

It just happened again

#8643 (record labels read through a coercion reported dead) has the same shape: a label reached through an aliasing relationship is not credited to the original declaration. The fix in #8645 was written against the batch pass first and would have shipped batch-only — the reactive half was added only because a reviewer asked whether the editor path was covered. Without that question this issue would have had a second entry.

Smaller things noticed nearby

• Reactive_type_deps.all_type_refs is only same_path_refs, so add_to_refs_builder would drop the impl↔intf refs that all_type_refs_from includes. Harmless today: both add_to_refs_builder helpers are dead code, never called.

What would actually prevent the next one

1. CI checks the two agree. Run the corpus in both modes, diff the finding sets. Cheap, and it would have failed the day #8217 landed. It fails today for the 7 above, so it lands with the re-export fix.
2. Delete the duplication. The genuinely mode-specific part is narrow — how declarations are indexed and how an edge is recorded. The rules themselves (which declarations participate, how they pair, which direction the edge runs) can be shared; #8645 does this for coercions with Dead_type.pair_coercion_labels / record_label_of_decl. Applied to re-export linking too, adding a rule to one pass and not the other stops being possible.
3. Or keep one path. If reactive is what ships, batch is a reference implementation that no user runs and that CI mistakes for the product. Retiring it removes the failure mode outright — but it wants (1) first, so the two are known to agree before one is deleted.