The optimizer twin is already cut — what §3.3 actually describes, and what duplication is left

Date: 2026-10-01 · Lane: “вырезать двойника оптимизатора” · Branch: wt/dedup-opt

The headline: the mandate’s premise is false, and deleting routing.rs would break the trainer

The task was: crates/dormouse-core/src/routing.rs is dead, cut it.

It is not dead. It is the live runtime path, and the twin it was supposed to replace was deleted four days before this lane was written, by 831e3a0 (2026-09-28, fix(train): the optimizer policy was declared twice and the two copies disagreed).

The trace, from the trainer entry point:

crates/dormouse-train/src/lib.rs:886   build_optim(&model, &cfg)
crates/dormouse-train/src/optim.rs:528   -> build_optim_mode(model, cfg, &cfg.opt)
crates/dormouse-train/src/optim.rs:480   -> optimizer_groups(model, cfg)
crates/dormouse-train/src/optim.rs:352   -> routing::routing(model, cfg.factors_fallback)   <== routing.rs RUNS

optim.rs no longer decides anything; it assembles what routing.rs declares, into burn::module::ParamGroups by ParamId, then verifies the install against the live tree (optim.rs:373 check_installed). The string policy is gone from the tree entirely — every symbol 831e3a0 deleted is absent:

$ grep -rn "MUON_PATH_MARKERS\|is_muon_param\|is_qk_param\|is_engram_table_param\|QK_HEAD_MARKERS\|ENGRAM_TABLE_MARKER\|effective_muon_markers" --include=*.rs crates/
crates/dormouse-core/src/routing.rs:426:  // (is_muon_param said the weight was Muon+, ...

One hit, and it is a comment remembering the divergence, not code. The markers are not in optim.rs any more; §3.3’s “builds the optimizer from path-string markers and is what runs” describes a file that no longer exists.

So the mandate’s condition — «ЕСЛИ тесты routing.rs проверяют что-то, чего optim.rs не проверяет — перенеси ЭТИ проверки, потом удаляй файл» — never fires. The tests of routing.rs and the checks in optim.rs are not two views of one policy; they are two different questions about one policy:

	asks	fails when
routing.rs Routing::check (:205)	is the DECLARATION total?	a parameter is in no group, or in two
optim.rs check_installed (:373)	does the INSTALL reproduce the declaration?	the ParamGroups the optimizer holds disagree with it

Deleting routing.rs would delete the declaration and take optimizer_groups, build_optim, the startup banner (lib.rs:899) and validate_routing with it. Not done. Not proposed.

What §3.3 got wrong, and the commit that made it wrong

The bullet is stale in a way that is worse than merely out of date: it names the wrong file as authoritative, and it was written after the fix.

routing.rs:265-268 (added d81d920, 2026-10-01, three days after 831e3a0):

/// Note this trait is the ID-based DECLARATION. The policy that actually runs
/// is the path-marker implementation in `dormouse-train/src/optim.rs`, and
/// they agree today; until that is reconciled, say which one you mean
/// (AGENTS.md §3.3).

git log -L puts that text in d81d920 (2026-10-01). 831e3a0 is 2026-09-28. A docs commit reintroduced the twin into the narrative after the code commit removed it — §1.4’s rule in its purest form: a claim that names no evidence, contradicted by the code in the same repo.

The same false claim is live in three more places, all still asserting a twin:

• docs/glossary.md:436-446 — a whole glossary section, “markers (path-string routing, the live implementation)”, documenting MUON_PATH_MARKERS, is_muon_param, is_qk_param as the running policy with citations to optim.rs:84,96,142,101,123,145. Every one of those line numbers is dead.
• docs/glossary.md:713 — disagreement row 16, “routing policy”, OPEN.
• docs/papers/muon-plus.md:186 — D17, “this is the id-based one, tests-only per routing.rs, with the path-marker copy deleted”. Half right: the copy is deleted, and then it calls routing.rs tests-only.

This is why AGENTS.md §3.3 is worth more than the two sentences it spends: a glossary section teaches a reader a vocabulary that does not exist, and docs/glossary.md is the file §1.7 makes binding (“If a document and the code disagree about a name or a default, fix one of them; never invent a third name”).

What duplication IS left: two, both small and both real

1. GroupCounts, declared twice — the real twin, ~15 lines

Same four fields, same meaning, two crates:

• routing.rs:65 — #[derive(Default, Debug, PartialEq, Eq)], four documented pub fields.
• optim.rs:440 — #[derive(Default, Debug)], same four, undocumented.

optim.rs’s is pub and re-exported at lib.rs:27, so it is API. The declaration is what the banner prints (lib.rs:899), so routing.rs’s is the one with the field docs that belong there.

Cut: delete optim.rs’s, re-export core’s.

2. Two module visitors over the same tree — IdCollector vs PathCollector

Both walk the live DormouseModel with a Vec<String> path stack and collect (path, id). optim.rs’s also collects the rank D.

• routing.rs:127 IdCollector — Vec<(String, ParamId)>, used at :161 (rest_of) and :206 (check).
• optim.rs:450 PathCollector — Vec<(String, ParamId, usize)>, plus param_paths (:471), pub and used by tests in lib.rs and tests/routing_policy.rs.

Identical stack discipline, identical join, identical visit_float. The rank is the only difference, and one impl costs one extra tuple slot.

Cut: routing.rs’s IdCollector is the wrong direction — it is in the lower crate and does less. param_paths already carries the rank, so core should own the one visitor and the train crate re-export it.

What is NOT a twin (checked, so nobody re-litigates it)

• Routing::check vs check_installed — declaration vs install (§ table above). Both load-bearing; optim.rs:355 calls check, then :357 calls check_installed.
• Group::ALL (routing.rs:54) — no caller. Dead, but it is a compile-time tripwire on the enum, not a policy copy; deleting it is a separate call and not this lane’s.
• role_of/Role — the what is this axis, used by future_byte.rs:81 in prose and by group_of. Not an optimizer group.
• Muon vs HeadWiseMuon (optim.rs:168) — two Optimizer impls serving two different groups, both counted on the eval line. Not a policy copy.

Line count

Net −~30 (a GroupCounts struct + a ModuleVisitor impl + one re-export), not the −200 the mandate projected: that projection assumed a 557-line file could be deleted, and it cannot be. The 557 lines of routing.rs are the live policy, and the tests in it are the ones that make a new arm failing to compile instead of silently training on AdamW.

Follow-ups (not this lane)

1. Group::ALL has no caller — routing.rs:54. Either use it in validate_routing (it would then assert every group is servable) or delete it. 3 lines either way; needs an owner decision on whether the tripwire is wanted.
2. docs/adr/0017-dormouse-fused.md describes the id-based declaration as the mechanism. That is now correct and no ADR change is needed — but the ADR should say the trainer calls it, since the “two implementations” framing came from here.