feat(parsers-v2): add Muse Glimmer unified and streaming tool-call parsers

[KrishnanPrash]

Overview

Adds Muse-Glimmer-30B parser support as a v2 UnifiedParser: one state machine that emits reasoning, content, and tool-call deltas in model order, plus the tool-only projection of the same scanner. The model uses ATEM tool markup and recipient-routed reasoning.

What this adds

• parsers/v2/src/unified/muse_glimmer.rs: the unified parser.
• parsers/v2/src/tool_calling/muse_glimmer.rs: MuseChannelScanner (recipient-dispatched) and MuseGlimmerToolStreamParser, the tool-only projection of the same scanner. Registered as muse_glimmer in create_tool_parser_for_family and create_unified_parser_for_family.
• Muse cases in the v2 golden corpus (the 33 shared scenarios plus the reasoning and tool interleaving cases).

Why a muse-specific scanner

WrappedBlockScanner dispatches on fixed marker literals. Muse decides the channel from the dynamic to=<recipient> header (the same <|start|> ... <|message|> opener begins reasoning, content, or a tool block), so there is no marker pair to declare. MuseChannelScanner owns the routing and reuses the shared primitives; one scanner drives both surfaces.

Validation

• cargo test --workspace pass, clippy -D warnings and fmt clean, check.sh ci green, check_family_coverage --family muse_glimmer 0/0, golden corpus regenerates byte-identical.
• v2 batch equals stream at every byte split; a real trace ships as a fixture.
• The v2 tool-call output matched the reference over the 112-record e2e suite with zero marker leaks. The full unified served path (reasoning plus tools through one parser) lands with the companion Dynamo change below.

Known limitations (documented and pinned)

• A marker split across a <|eom|> seam can survive one strip pass and reappear in assembled text. Requires model-invalid input (a reserved-token fragment as prose across a real control token). Pinned as a chunk-invariance exception.
• Integer arguments past u64::MAX format via f64.
• <|start|> inside an ATEM parameter value is read as a channel switch.

Companion Dynamo change (separate PR)

To serve muse through v2, Dynamo consumes the unified parser (create_unified_parser_for_family) for muse by default, which replaces both the v1 reasoning parser and the tool jail. Dynamo's parser_requires_special_tokens must also list the muse aliases, or serving strips the channel framing before any parser runs (Inkling precedent, dynamo #11823).

Versions

• dynamo-parsers-v2 0.2.0 to 0.2.1 (additive, on top of the refactor(parsers): align the unified parser surface with peer traits, and let vendors supply their own #178 unified-surface refactor).

[github-actions]

📊 Conformance matrix rendered — view in CI summary

[indrajit96]

LGTM!

Core parser behavior and targeted tests look solid.
Some Nits and versioning questions

[keivenchang]

Thanks.

[keivenchang]

hey Dallas, nice work on this. digging into the conformance table after approving — muse_glimmer only covers 36 of 81 unified cases. guided-decoding and prefilled-response groups get skipped since there's no initialize_request override yet, so the generator would panic if it tried. not asking you to fix this one, it's already documented as a known gap. can you add a tracking issue (or TODO) for the missing guided-decoding/prefilled-response coverage so it doesn't get lost?

[keivenchang]

conformance (tests) to be completed:

[keivenchang]

Just gonna pre-approve to unblock you.

[KrishnanPrash]

Added the TODO for the guided-decoding/prefilled-response coverage. Thanks for the approval @keivenchang!