feat(tfmx): play single-file Hippel TFMX and Future Composer formats

[indigo423]

Why

The vendored libtfmx backend already decodes single-file Amiga formats — Jochen Hippel TFMX/MCMD (.hip .hipc .hip7 .mcmd) and Future Composer (.fc .fc3 .fc4 .fc13 .fc14 .smod) — but the app only ever handed it two-file Huelsbeck pairs, so those files were rejected at every layer. This addresses issue #89, where the library author (mschwendt) asked "why do you reject the other file types? Like Jochen Hippel's TFMX files? Like Future Composer? MCMD?" — and the maintainer agreed we shouldn't.

Scope note: issue #89 also reports a Huelsbeck pair failing to play on prod — that's a separate bug (needs a repro file), out of scope here.

What changes

Single-file formats now play from Library browse/search and Local drop, alongside the existing pairs:

• Allowlist TFMX_SINGLE_EXTENSIONS + isTfmxSingleFile/tfmxSingleExt. Deliberately kept out of isModuleFile (distinct engine) and excluding the ambiguous .mdat/.tfm/.tfmx names — a lone .mdat could be an orphan Huelsbeck half or a complete Hippel file, so it stays pair-only (deferred).
• New Source variants tfmx-single-local / tfmx-single-library (single buffer, no sample half), threaded through getBuffer, the play() dispatch, and the drawer picked-state.
• Worklet loads a lone music-data file: skips the .sam write and names the MEMFS file with the real extension (e.g. song.fc) so libtfmx's sidecar-guessing doesn't hunt for a phantom .sam; tfx_load then content-detects the format.
• Server: the byte-server perimeter allows single-file extensions with no partner check (orphan-half rejection preserved for pair formats); listing + search surface singles; tfmx-random left pair-only.

No vendored-library or WASM change — the engine already plays these formats.

Verification

Driven end-to-end against a running server (via scripts/run-driver.mjs):

• /api/library/file serves .fc/.hipc without a partner → 200; orphan .tfx/.mdat still → 404 (perimeter intact); tfmx-random → 404 (singles excluded).
• A Future Composer and a Hippel file decode and play through the tfmx-single-library engine path (screenshot: "Playing from: tfmx-single-library").
• make verify green (lint 0 errors, typecheck, 232 unit tests incl. 16 new, build).

Notes / follow-ups

• Single-file rows render the filename (e.g. chambers-balance.fc) rather than a synthesized (FC)/(Hippel) label — the extension carries the format identity. Easy to add a family tag if wanted.
• Perimeter behavior verified live (curl); no automated handler integration test yet.
• Adds scripts/run-driver.mjs, a Playwright driver for headless player smoke-tests (documented in the local run-coolmodfiles skill). Heads-up: TFMX decode fails under Playwright's bundled Chromium because its AudioWorkletGlobalScope lacks TextDecoder — an env gap, not a product bug; real browsers have it.

OpenSpec change: add-tfmx-single-file-formats (gitignored).

🤖 Generated with Claude Code

[indigo423]

Applied max-effort code-review follow-ups in abb0430:

A — single-file auto-advance / error recovery (the reachable one): pickRandomNext returned null for both tfmx-single-* arms, so a getBuffer fetch failure on a user-picked single teleported to a random Mod Archive track — contradicting the anti-teleport comment right above the fallback, and inconsistent with library/tfmx-library (which retry in-family). Now tfmx-single-local mirrors tfmx-local (random-pick from the dropped set) and tfmx-single-library does a folder walk (pickNextInFolder now understands singles[]) with an /api/library/random fallback — staying in the library.

B — single-vs-pair-half precedence: detection classified by extension alone, so a prefix-Amiga pair whose base ends in a single-file token (mdat.fc + smpl.fc) was mis-split into bogus singles on local drop and double-listed by /api/library. detectSinglesInDir, the local detectTfmxPairs, and the search single-arm now all require parseHalfName()===null — pair detection takes precedence uniformly (matching what search.ts already did).

Verified live: mdat.fc+smpl.fc lists as one pair (not singles); genuine .fc/.hipc still list + serve; orphan mdat.fc → 404. +2 regression tests (234 passing); typecheck + lint clean.

Not changed: the latent worklet musicExt .tfx fallback (finding #5) — kept as-is since it's unreachable today; can harden separately if desired.