Make explicit-file changes, stage, and commit truly targeted

[c0da-ricon]

Problem

Explicit-file Notes workflows still perform whole-corpus change detection before applying the requested filename filter. On a large deobfuscated repo, the supposedly targeted path is too slow for interactive editing.

Observed with shiv:notes 0.8.19 in ricon-family/fold on 2026-07-17:

• 437 manifest entries / 438 readable Markdown files;
• git status --short --branch: 0.062s;
• notes parse notes/notes-performance-fast-path.md: 0.804s;
• notes changes notes/notes-performance-fast-path.md: exceeded a 20s hard timeout;
• notes changes --summary: exceeded a 20s hard timeout;
• targeted notes commit --dry-run ... notes/notes-performance-fast-path.md: exceeded a 20s hard timeout.

Each timeout terminated the command. No result was treated as complete.

Source finding

In 0.8.19:

• .mise/tasks/changes calls detect_changes "$abs_notes_dir" before filtering by ARGS;
• .mise/tasks/stage does the same before selecting requested paths;
• .mise/tasks/commit delegates to stage, then calls whole-corpus detection again during post-commit verification.

A component probe on the same Fold checkout found:

• stale-readable detection: about 3s;
• batched lookup of 437 HEAD objects: 0.232s;
• two batched attribute scans: 0.106s;
• batched hashing of all 437 readable notes through configured Git filters: about 14s.

The explicit filename currently improves safety and scope, but not discovery cost.

Desired direction

Make explicit-file operations genuinely targeted while preserving fail-closed safety for the selected notes.

--all and deep diagnostic paths may remain whole-corpus operations. Ordinary edits to one or a few known notes should not hash every readable note multiple times.

Acceptance

• notes changes <known-file> compares only the requested note and returns its readable diff/summary.
• notes stage <known-file> validates and stages the requested note without first detecting every unrelated change.
• notes commit <known-files> does not perform repeated whole-corpus scans before and after the commit.
• Selected-note guards still cover stale readables, dual-present conflicts, double tracking, manifest mapping, hook availability, and post-commit correctness.
• --all retains complete whole-corpus behavior.
• Large-corpus regression coverage demonstrates that targeted cost scales with selected files rather than total notes.

Current workaround

Batch related note edits and run one supported targeted notes commit at a coherent boundary. Avoid repeated notes status / notes changes rituals during editing. This reduces frequency but does not fix the underlying cost.

[c0da-ricon]

Additional real-path evidence from the same Fold checkout:

notes commit -m 'notes: add performance fast path' \
  notes/notes-performance-fast-path.md \
  notes/notes-managed-repo-workflow.md

The supported two-file commit completed successfully in 152.283 seconds. It staged two notes, auto-obfuscated them, committed, deobfuscated, and verified clean state.

This was one deliberately batched commit at a coherent boundary. The result confirms that the performance problem is not limited to dry-run or a timeout wrapper; the ordinary successful targeted path is far outside an interactive editing budget.