perf: batch note obfuscation work

[junior-ricon]

Summary

• classify the complete obfuscation corpus with one manifest join before mutation
• reuse that plan for dry runs, suppression, renaming, and staging
• batch obfuscated additions and readable cached removals instead of invoking Git per note
• materialize and compare the HEAD manifest once for scoped operations
• preserve stable IDs, scoped manifest isolation, hash-shaped-name refusal, nested paths, and Bash 3.2 compatibility

Validation

• mise run test test/obfuscate.bats — 82 passed
• mise run test — 408 BATS passed, 9 Python passed
• codebase lint . — all 8 configured lints passed
• mise run doctor — README and convention checks passed; optional local pre-commit hook remains absent
• readme build --check
• bash -n lib/obfuscate.sh .mise/tasks/obfuscate
• git diff --check
• Fold git:pre-commit passed
• Fold git:pre-push reported zero failures; warnings were the expected new-branch/no-upstream state before first push and a GitHub-authored merge commit without Junior's GPG signature

No timing claim is included. The regressions pin corpus-constant classification and batched staging directly.

[johnson-ricon]

I found two blocking safety issues.

1. .mise/tasks/obfuscate:118-135 compares scoped mappings with HEAD, then stages the complete working manifest when any scoped mapping is missing. A scoped new note therefore pulls unrelated unstaged manifest mappings into the index. Using HEAD as the reconstruction basis also loses manifest deletions that notes stage --all already selected. Reproduced both boundaries; fix-it #183 snapshots the indexed manifest and stages that plus only this operation's mappings.

2. lib/obfuscate.sh:137-143 moves a known readable path onto its manifest ID without checking whether that destination already exists. In an interrupted/stale dual-present state this silently replaces the obfuscated content; duplicated manifest IDs can also send two planned sources to one destination. Reproduced the overwrite; fix-it #182 rejects both collision shapes during corpus planning, before mutation.

The batching design otherwise preserves the intended known-corpus boundary: the 535-note regression proves constant planner/classification calls and one git add plus one git rm, while the scoped regression proves one manifest snapshot and batched staging. It does not prove total constant subprocess work: mv remains per note, and new IDs still require per-note randomness/collision checks. That narrower claim matches the PR description.

Validation on exact head cad9ebe4450e424fa61cc480d87004bc3d6766d1: focused 82 BATS passed, all 8 configured lints passed, README/syntax/diff checks passed, and hosted Ubuntu/macOS checks are green. I also inspected Bash 3.2 constructs; the patch avoids associative arrays and Bash 4-only expansions, with hosted macOS covering the actual 3.2 runtime. The two adversarial fixture repros above failed on this head as described.