docs(plan): score engraving and polyphony roadmap

[akshayswaroop]

Summary

Documentation-only roadmap for improving ExactKeys piano engraving while preserving the repository's fail-closed accuracy contract.

This PR contains no runtime implementation and changes no certification behavior.

Planned areas

• Support independent same-staff voices using an explicit score voice model and valid MusicXML <backup> / <forward> timelines.
• Preserve the existing exact-onset/equal-duration chord behavior while extending only polyphony that can round-trip without dropped or altered events.
• Investigate rolling key-change detection; all proposed window and confidence values are hypotheses requiring fixture-backed validation.
• Add beat-group-aware duration spelling while preserving exact tied-note recovery.
• Require production-path adversarial tests and 100% MusicXML reparse fidelity for every future certifying change.

Contract boundaries

• score-note-v1 remains MIDI-relative and fail-closed.
• Audio-derived notes remain permanently uncertified under the current contract.
• Voice allocation and inferred key changes are notation decisions, not unqualified accuracy claims.
• Unsupported polyphony continues to abstain until exact representation is proven.

Scope

Exactly one file:

• docs/plans/2026-08-13-score-engraving-and-polyphony-roadmap.md

Verification

• Rebuilt directly from current main; no PR feat(engine): add key estimation, multi-voice staff splitting, Web Worker inference, and CLI audio pipeline #3 commits remain.
• git diff --check origin/main...HEAD passes.
• PR diff contains one commit and one documentation file.
• Hygiene scan found no local paths, private fixture names, credentials, or unsupported empirical claims.