feat(stores): local-fs and user-local outside-repo backends (feature 034)

[mlieberman85]

Summary

Two new filesystem-backed Store backends inside darnit-core that write outside the audited repository:

• local-fs: config-driven root (absolute, ~-expanded, or $VAR-templated). Fail-fast on missing env vars.
• user-local: platform-conventional root (XDG on Linux, Apple support/cache dirs on macOS, %LOCALAPPDATA% on Windows). No root config needed.

Both are selectable via .baseline.toml:

```toml
[stores.attestation]
backend = "local-fs"
root = "$DARNIT_ATT_ROOT"

[stores.report]
backend = "user-local"
```

Motivation

Feature 033 (#396) shipped the pluggable-stores abstraction. Its filesystem defaults all write inside the audited repo (<repo>/.darnit/attestations/, etc.). This PR closes the "put it somewhere else on the local filesystem" gap that OSPO leaders keep hitting (consolidate attestations across 30 repos into one directory) and that CI operators keep hitting (cache in the runner cache dir, reports in the runner artifacts dir).

Motivating scenarios (in spec.md):

1. OSPO leader consolidates every audit's attestation into ~/darnit-attestations/.
2. CI operator points cache at $RUNNER_CACHE_DIR/darnit and reports at $RUNNER_ARTIFACTS_DIR/darnit-reports.
3. Individual dev uses backend = "user-local" and lets XDG / macOS conventions decide.

Multi-repo templating: env-var interpolation is the escape hatch. root = \"\$DARNIT_ATT_ROOT\" in every repo's .baseline.toml plus one per-machine env var. No new config layer (spec Q1 clarify) -- this stays within .baseline.toml.

Design highlights

• No new runtime dependencies (FR-014). platform_paths.py hand-rolls XDG / Apple / Windows conventions in ~90 lines rather than pulling in platformdirs.
• No new Store Protocol methods. Both backends satisfy the feature-033 Protocols by delegating to the in-repo classes after root resolution.
• .project/project.yaml stays in the repo (FR-009). Neither backend is registered under darnit.stores.project; [stores.project] backend = "user-local" raises `StoreNotInstalled` at `resolve_stores` time before any control runs.
• Path traversal impossible by construction (SC-005). Shared _sanitize_filename regex from feature 033 handles it.
• Cross-filesystem cache writes safe (research R-004). The delegate already writes the tempfile into the target's directory.
• Feature 033's zero-config test passes unchanged (SC-003). New test_us4_zero_config_local.py extends the invariant to the new backends.
• Loud error surface (SC-008). LocalFs*/UserLocal* wrap the delegate's `StoreOperationError` with a message that names the backend, kind, and resolved path.
• Info-log on every outside-repo write (FR-015). darnit.stores.local logger emits wrote <kind> (<backend>): <resolved-path> per artifact. Pre-feature filesystem defaults emit zero lines to that logger.

Constitution check

All five principles PASS (from plan.md):

Principle	Status
I. Plugin Separation	PASS (new code in darnit-core; no cross-package imports added)
II. Conservative-by-Default	PASS (no silent fallback; per-Protocol failure semantics preserved)
III. TOML-First Architecture	PASS (config surface stays TOML)
IV. Never Guess User Values	N/A (storage backends don't produce user-judgment values)
V. Sieve Pipeline Integrity	N/A (backends are not sieve passes)

Deferrals (documented in tasks.md)

• T015: LocalFsProjectStateStore deferred. Would require reimplementing .project/-prefix-free YAML I/O; better handled in a separate PR alongside the audit driver's use of bundle.project.
• T041: Windows CI runner integration test. Deferred because no Windows CI runner is currently configured. Windows path resolution is unit-tested via platform.system() monkeypatch, so SC-004 is satisfied at the unit level.

Spec artifacts

Under specs/034-local-output-store/:

• spec.md: 4 user stories, 15 FRs, 9 SCs, 3-question clarify session
• plan.md, research.md, data-model.md, contracts/, quickstart.md
• tasks.md: 41 tasks in 7 phases, 39/41 completed (T015, T041 deferred with rationale)
• checklists/requirements.md: 16/16 pass

Test plan

• Baseline before change: 97 stores tests pass.
• After Phase 3 (US1 MVP): 120 pass.
• After Phase 4 (US2): 135 pass.
• After Phase 5 (US3): full stores suite 157 pass.
• Feature 033's test_us2_zero_config.py passes unchanged (T032).
• Full workspace sweep: 3061 pass, 26 skip, 0 fail.
• ruff check . clean repo-wide.
• validate_sync.py clean.
• Structure guard: all changes within stores/defaults/, pyproject.toml, stores/__init__.py, docs/plugin-authoring/stores.md, tests/darnit/stores/ (plus speckit bookkeeping).
• FR-014 no-new-deps guard: pyproject.toml diff is entry-point registrations only.

Files changed

Production code (new):

• packages/darnit/src/darnit/stores/defaults/local_fs.py (~200 lines)
• packages/darnit/src/darnit/stores/defaults/user_local.py (~120 lines)
• packages/darnit/src/darnit/stores/defaults/platform_paths.py (~90 lines)

Production code (modified):

• packages/darnit/src/darnit/stores/defaults/__init__.py (re-exports)
• packages/darnit/pyproject.toml (6 entry-point registrations)

Docs:

• docs/plugin-authoring/stores.md ("Writing artifacts outside the repo" section)

Tests (new):

• tests/darnit/stores/test_local_fs_backend.py
• tests/darnit/stores/test_local_fs_helpers.py
• tests/darnit/stores/test_local_fs_isolation.py
• tests/darnit/stores/test_local_fs_logging.py
• tests/darnit/stores/test_platform_paths.py
• tests/darnit/stores/test_us4_zero_config_local.py
• tests/darnit/stores/test_user_local_backend.py

Speckit artifacts: specs/034-local-output-store/ + speckit-bookkeeping updates to .specify/feature.json and CLAUDE.md.

[Marc-cn]

Tested on Linux. Stores suite 157/157. Both backends behave as described. Nothing landed inside the audited repo.
Zero-config check: audit on tqdm is byte-identical to main across all 66 controls, so SC-003 holds. lgtm