[Product] Add an active rehearsal player with precise looping and role controls

[seonghobae]

Parent: #958

Buyer-visible gap

BandScope currently concentrates on importing and explaining analysis. The buyer still needs a product surface that turns that analysis into repeatable rehearsal action. Static cue cards, timelines and exports do not by themselves let a player practise a difficult section, count in, loop it, adjust speed, mute/solo a reference part, and verify the next entrance.

This issue completes the core product loop without turning BandScope into a DAW.

Product outcome

A musician selects any detected or user-defined section and rehearses it immediately with a deterministic local transport:

select section or cue
→ hear count-in
→ loop exact passage
→ slow down or restore tempo
→ solo/mute available reference stems
→ navigate cue-to-cue by keyboard
→ preserve progress and return to the same rehearsal state

The analysis remains advisory. The player must expose confidence and allow users to correct loop boundaries when detection is wrong.

Required scope

Transport

• Play, pause, stop, seek and restart from the current rehearsal cue.
• Sample-accurate or explicitly bounded loop start/end semantics across supported formats and sample rates.
• Count-in with configurable beats and meter-aware defaults where meter is available.
• Playback-rate control with a declared quality boundary; preserve pitch when the chosen backend supports it and label fallback behavior honestly.
• Keyboard-accessible transport and focus-safe shortcuts that do not trigger while typing in text fields.
• A single authoritative transport state machine; do not let UI controls, native playback and project persistence race independently.

Rehearsal controls

• Choose a detected section, timeline range, action cue or manually corrected range.
• Move to previous/next cue without losing the active loop.
• Persist per-project last position, active loop, playback rate, count-in and selected role.
• Solo/mute only when a licensed/locally generated source actually exists; never fabricate isolated stems or imply separation success.
• Make entry, dropout, stop, pickup and role-range cues available from the same transport surface rather than as disconnected micro-features.
• Integrate existing product slices such as feat(workspace): loop tonight's first section on the map #903, feat(workspace): open tonight's first notes on the groove map #905, feat(workspace): guide tonight's first entrance on map and player #912, feat(workspace): guide tonight's first dropout on map and player #914, feat(workspace): guide tonight's first pickup on map and player #916, feat(workspace): guide tonight's first stop on map and player #934, feat(workspace): guide tonight's first intro on map and player #943, feat(workspace): guide tonight's first bridge on map and player #946, feat(workspace): guide tonight's first verse on map and player #947, feat(workspace): guide tonight's first pre-chorus on map and player #955 and feat(workspace): name tonight's first playable range on the map #957 through one coherent player contract after their dependency/ownership state is reconciled.

Runtime architecture

• Prefer a Rust-owned playback/transport core for timing, bounded local file access and cross-platform lifecycle; keep the React layer as a typed control/view surface.
• The real-time audio callback must not perform unbounded allocation, blocking I/O, network access or lock-heavy work.
• Decode only after the resource-admission checks from [Security] Enforce one canonical local-audio resource budget before analysis #781/fix(audio): establish canonical local-audio resource policy #866 and support cancellation before expensive decode or analysis.
• Define explicit behavior for unsupported codecs, missing files, replaced files, device changes, suspended output devices and project close during playback.

Accessible and honest UI

• All controls expose name, role, state, shortcut and current value to assistive technology.
• Waveform/timeline visual selection has an equivalent numeric start/end editor and exact-value table.
• Show when a loop came from detected analysis versus a user correction.
• Never report playback success merely because a command returned; confirm observable transport state and rendered position.

Acceptance criteria

• A licensed local demo and a user-selected local audio file can be played, paused, sought and stopped through the production desktop boundary.
• Loop timing error is measured against deterministic PCM fixtures at multiple sample rates and remains within a documented bound.
• Count-in, loop restart, playback-rate changes and cue navigation are deterministic under repeated tests.
• Device loss, file replacement, unsupported codec, cancellation, app backgrounding and project close fail safely without leaking paths or hanging the UI.
• Transport state restores after ordinary restart and coordinates with the crash-safe project/autosave issue under [Product Readiness] Complete BandScope 1.0 with measurable rehearsal outcomes #958.
• Keyboard-only and screen-reader journeys can select a section, start a count-in, loop, pause and clear the loop.
• Existing cue/action PRs are integrated through one public contract or explicitly superseded; no duplicate local transport stores remain.
• Production statement/branch coverage and public API documentation are 100% for repository-owned player code.
• CPU use, memory, underrun/dropout count and worst-case control latency are measured on the supported hardware profiles.
• Current-head checks, security gates, independent approval and zero unresolved threads pass.

Real-world tests

• 44.1/48/96 kHz, mono/stereo and short/long files;
• pickup before bar one, odd meter, tempo change and silence around a boundary;
• repeated start/stop and rapid loop-boundary edits;
• Bluetooth/USB device removal and default-device switch;
• very short loops, invalid reversed ranges and range beyond decoded duration;
• pause during count-in and stop during loop restart;
• user-corrected section boundary retained across restart;
• stem unavailable despite a role recommendation;
• keyboard focus inside text/numeric input;
• CPU pressure without UI deadlock or unbounded queue growth.

Non-goals

• No multitrack recording, plugin hosting, automation lanes or free-form mixing console.
• No cloud streaming requirement.
• No claim of source separation quality beyond evidence from [Product Gap] Add real-audio MIR accuracy acceptance benchmarks #770.
• No arbitrary filesystem or generic media execution capability.