feat: Make turbo watch Cargo-aware (#13280)
## Why
Next in the Cargo workspace series. `turbo watch` in a Cargo-enabled
repo has two failure modes today: the watcher's private package graph is
JS-only (crate edits map to nothing), and nothing protects against the
feedback loop where Cargo's continuous `target/` writes re-trigger the
very tasks that produced them. Per the series design, the fix is a trait
surface — `Toolchain::watch_spec` — with both implementations stated
honestly.
## What
- `Toolchain::watch_spec`: workspace-definition file names/paths (change
→ full rediscovery) and build-byproduct directory prefixes (events
dropped). Plain-string data per the toolchain design rules
- **Cargo**: any `Cargo.toml` or the root `Cargo.lock` triggers
rediscovery (the crate set or its edges may have changed); manifests
under `target/` are exempt; all events under root `target/` are dropped
— the feedback loop must not depend on a `.gitignore` entry
- **JavaScript**: declares nothing extra, documented as the real answer
— JS workspace redefinition is already caught by the change mapper's
conservative all-packages fallback
- The watcher builds its package graph with the same toolchains a run
registers (plumbed from the future flag in `watch.rs` and `daemon.rs`),
so watch sees the run's package set
- Known gap documented in ARCHITECTURE.md: hash-watcher content dedup is
JS-glob-based, so a no-op crate save re-runs as a fast cache hit
- Rider: one-line collapsible-if fix in `turborepo-napi` that currently
breaks `cargo lint` on main (and therefore every contributor's pre-push
hook) — flagged here rather than a separate PR since it blocks this
branch's own push
## How
Flag off: watcher behavior is unchanged (the merged spec is empty;
classify short-circuits identically). Flag on, verified live under
`turbo watch` on the `cargo_monorepo` fixture: dependency-crate source
edit re-runs the entrypoint; JS edit re-runs only the JS package; writes
under `target/` (including a `Cargo.toml` cargo places there) trigger
nothing; adding a new crate manifest triggers full rediscovery. Unit
coverage: three classify tests (manifest → rediscovery, lockfile →
rediscovery, target/ suppression) each asserting the flag-off behavior
too.
