feat(overlay): add the Lore capture overlay MVP (verified core + Tauri shell scaffold)

[tcballard]

Summary

A first MVP of the capture overlay (Host B of lore-capture-surfaces,
architecture in lore-capture-overlay). It is a lore-* product (ADR-068)
developed here as a staging directory (lore-overlay/) to extract to its own
itsthelore/lore-overlay repo later — the develop-in-repo-then-extract pattern
used for the VS Code extension.

Two surfaces:

• lore-overlay/core/ — the platform-agnostic capture "brain" (Rust lib).
  Fully tested here.
• lore-overlay/app/ — the macOS-first Tauri v2 desktop shell. Authored
  scaffold; not built/run (it targets macOS; this was developed in a Linux
  container).

The product name is provisional ("overlay" is a working title), per the
maintainer's note.

Scope

Included

• The capture core: the two-gate flow (ADR-077) over three trait seams (rac
  engine / model gateway / GitHub). propose (Gate-1 fidelity) → publish
  (writes, validates with rac, opens a draft PR only). The Publisher
  trait has no approve/merge method, and publish refuses any non-draft PR —
  a host on this core cannot self-approve.
• A bring-your-own OpenAI-compatible gateway (ADR-035); the engine seam shells to
  rac (ADR-063); no AI in the engine (ADR-002/067).
• The Tauri shell scaffold (tray + global hotkey → modal → the two commands) and
  a minimal two-gate frontend.
• A README.md that states exactly what is verified vs not.

Excluded

• The desktop app is not built, run, signed, or notarized — that needs macOS.
• Real settings persistence + OS-keychain secret storage (scaffolded via env vars).
• Desktop GitHub-App device-flow auth (the core takes a bearer token; acquiring it
  is the shell's job — an open question in the design).
• No engine, corpus (rac/), or Python change.

Product / Architecture Decisions

• Core in Rust so it is the Tauri backend, yet cargo test runs on Linux
  without the macOS webview. Network clients (reqwest) sit behind a net feature
  so the test suite is fast and offline.
• The two gates are encoded in the type system: separate propose/publish
  steps, and a Publisher trait with no approve path.

Verification

Ran

• cd lore-overlay/core && cargo test — 4 hermetic tests pass (propose→publish
  keeps frontmatter + opens a draft PR; non-draft PR is refused; minted-id parse;
  schema/draft/fill/validate over fakes).
• LORE_TEST_RAC="…" cargo test real_rac_client_reads_schema_when_configured —
  the real RacClient shells to rac schema decision --json and parses it.
• cargo check --features net — the gateway + GitHub clients compile.
• cargo fmt --check — clean.

Not run (cannot, in a Linux container)

• cargo tauri build / dev, code signing, notarization — require macOS.

CI

• No rac/, src/, or tests/ change, so the Python/rac CI tier is unaffected;
  the Rust under lore-overlay/ is verified locally as above (CI does not build it).

Review Path

1. lore-overlay/README.md — what this is, verified vs not, how to build on a Mac.
2. lore-overlay/core/src/flow.rs — the two-gate orchestration.
3. lore-overlay/core/src/{rac,gateway,github}.rs — the three seams (traits + real impls).
4. lore-overlay/core/tests/flow_test.rs — the suite.
5. lore-overlay/app/src-tauri/src/lib.rs — the shell wiring (scaffold).

Notes For Reviewer

• This is a staging location; extraction to itsthelore/lore-overlay and the
  product rename are follow-ups.
• The shell's exact Tauri-plugin API calls are an authored scaffold (uncompiled
  here); treat them as a starting point to verify on macOS.

Implementation Process

Implemented with AI assistance; final scope, review, and acceptance decisions were
made by the maintainer.