refactor: introduce client-side VaultFs boundary

[enieuwy]

Goal

Introduce the client-side VaultFs storage boundary required before rebasing the headless CLI work. This PR migrates DiskMirror off direct app.vault.* file operations and onto a small filesystem interface, with an Obsidian adapter for the plugin and an in-memory adapter for tests.

Opening as a draft intentionally: please confirm this scoped PR-A boundary before detailed review.

Scope

Included:

• VaultFs contract, errors, stats/list/read/rename/write/delete operations.
• Shared vault path normalization (NFC, slash normalization, no leading/trailing slash).
• ObsidianVaultFs preserving Obsidian semantics (Vault.delete, FileManager.renameFile).
• FakeVaultFs and targeted DiskMirror tests for create/update/delete/rename, suppression fingerprints, frontmatter-blocked writes, target-exists rename, and missing-source rename.
• DiskMirror now uses VaultFs for disk reads/writes/deletes/renames.

Explicitly not included in this PR:

• The remaining raw app.vault.* callsites in src/main.ts for reconcile scans, vault event handlers, direct reads, and disk-index stat calls.
• CLI NodeVaultFs / NodeDiskMirror refactor. That is the PR-B rebase follow-up after this boundary lands.
• Orchestration unification into a shared DiskMirrorCore.

Follow-up after this merges

Rebase PR #16 (feat/headless-cli-fresh) on top of main, drop its duplicate normalizer, add NodeVaultFs, and move the CLI filesystem hardening behind that adapter while preserving the PR #16 hardening checklist.

Verification

• node --import jiti/register --test tests/sync/diskMirror-vaultFs.test.ts
• npm run build
• npm run test:regressions
• git diff --check
• git log --oneline upstream/main..HEAD shows one refactor commit

Not run locally:

• npm run test:e2e:obsidian failed before launching tests because wdio is not installed in this workspace (sh: wdio: command not found, exit 127). The two-peer manual Obsidian smoke was not run in this headless workspace for the same missing local harness reason.

[enieuwy]

Quick check before I rebase — upstream/main has moved ~120 commits since this branch was opened (runtime extraction into src/runtime/*, INV-SAFETY-02 in 48cfe67 + af61c03, the DiskMirror rewrite for RemoteDeleteDecision / PreservedUnresolvedRegistry / per-path write locks / setDiskWriteCallback, and the new canonical origin set in src/sync/origins.ts). Rebasing the existing DiskMirror migration on top of all that is doable but non-trivial, and the answer to two questions changes the shape:

1. Do you still want the client-side VaultFs seam as proposed? Text-only interface (normalize, readText, writeText, delete, rename, stat, listMarkdown) with ObsidianVaultFs / FakeVaultFs adapters, shared normalizeVaultPath (NFC + slash normalization), and exclude / frontmatter-guard behaviour staying where it currently lives. Yes / no / different shape?

2. If yes, do you want this PR shrunk to interface + adapters + tests only, with the DiskMirror migration as a focused follow-up? That makes the review trivial against current main and lets the recent DiskMirror rework settle. The alternative is to keep the current "boundary + DiskMirror migrated through it" shape, which is a more involved rebase against the new RemoteDeleteDecision / PreservedUnresolvedRegistry code paths.

Happy to push either shape on your call. I'll leave the draft as-is until I hear back so we don't churn on a direction you don't want.