Report node errors to Sentry, stamped with the build they came from

[bkearns]

ferrosa's half of the per-repo diagnostics work.

What was missing

There was a telemetry layer already, gated on FERROSA_TELEMETRY_ENABLED, but no path for an error to leave the machine. A node that logs a failure to a console nobody is attached to has not reported it.

What this adds

Errors become Sentry events when the installer configures FERROSA_SENTRY_DSN. WARN becomes a breadcrumb, not an event — context for the error that follows. Everything below WARN stays in the log: a storage engine at info level would exhaust the quota in minutes and hide every error underneath the noise.

The client starts before the subscriber. The layer is inert without it, and the errors most worth having from a database are the ones raised while it is still coming up. The guard is bound in main so it outlives the process; dropping it stops sending, which loses exactly the shutdown errors.

The disabled-telemetry branch moves from the fmt() builder to registry composition so the Sentry layer can sit beside the writer. Filter and writer unchanged.

Build identity

Events carry ferrosa@version+sha. build.rs prefers FERROSA_BUILD_SHA from CI, falls back to git rev-parse, and says unknown last — a stale SHA would be believed. A dirty tree is marked.

The release workflow should set FERROSA_BUILD_SHA so tagged builds carry the commit CI checked out.

The DSN is not in this repository

ferrosa is PUBLIC. A DSN is not a secret — it's a write-only ingest key meant to be embedded — but one committed to a public repo can be used by anyone to write into the project, and rotating it then needs a release. Empty (what an unset shell variable looks like) is absent, not broken; plain http is refused, since it would put diagnostics on the wire in the clear.

Dependency choice, deliberately

sentry 0.49, not 0.34. The older release pins rustls 0.22, whose rustls-webpki 0.102 carries four RUSTSEC advisories — cargo deny caught exactly this on the ferrosa-memory branch. One rustls-webpki in the lock now, the patched 0.103. rustls rather than native-tls because CI builds this for x86_64-unknown-linux-musl.

Adding a diagnostics dependency that ships known vulnerabilities would be a poor trade for observability.

Verification

• 7 tests in sentry_reporting
• cargo clippy --all-targets -- -D warnings and cargo fmt --check pass
• 1786 passed, 0 failures under the exact CI invocation

One note on the suite: a plain --all-features run also surfaces real_typed_edges_paged_scan_delivers_every_distinct_row, which fails loud without FERROSA_TEST_TYPED_EDGES_DIR by design. CI skips it explicitly, and so does the command above. I checked main is green before concluding that.

The __eh_frame linker warning on debug builds is pre-existing on main — I verified by stashing — and RUSTFLAGS isn't set in CI, so build warnings aren't errors there; clippy's -D warnings is the gate and it passes.