Split specialized chat handlers into focused per-format files

[leehack]

Context

PR #391 introduced a roughly 1,085-line internal file containing Kimi K3, MiniMax M1/M3, DeepSeek DSML, Muse Glimmer, Laguna, and shared parser/grammar helpers. This conflicts with the repository convention of focused handler files and makes correctness review harder.

Improvement

After the active P1 regression recovery lands, split each format handler into a focused file and retain only narrowly shared internal helpers.

Acceptance criteria

• One focused handler file per format family with clear internal ownership.
• Shared helpers remain only where duplication is materially reduced.
• No public API, routing, grammar, parser, streaming, or tool-choice behavior changes.
• No helper duplication or circular imports.
• Full template parity, VM/Chrome suites, coverage, analyzer, formatter, and representative real-model smoke remain identical.
• Perform this after Regression: DeepSeek V3.2 DSML function calls are not parsed #394-Regression: MiniMax M1 grammar omits JSON quotes around tool names #399, Regression: DeepSeek forced-open thinking swallows DSML tool calls #402, Regression: direct-Jinja required-tool grammars reject mandatory reasoning prefixes #406-Regression: MiniMax M3 rejects valid zero-argument tool calls #408, and Regression: specialized tag parsers lose schema-directed argument types #410 are resolved to avoid repair conflicts.

Priority

P2 maintainability follow-up; zero-regression recovery remains first.