Fix two latent scamp-go bugs: announce-parse panic + Send error swallow

[riccardo-perotti]

Both bugs were uncovered during the golangci-lint rollout (dead-assignment lint hits that were masking pre-existing bugs). Filed as ROP-9250.

Bug 1 — newServiceProxy panics on malformed announce data

serviceproxy.go: after unmarshaling the announce, classRecords[0..7] and (per action) actionsRawMessages[0..1] were indexed with no length check — short/malformed discovery data → index-out-of-range panic. Added two length guards (classRecords < 8, actionsRawMessages < 2) that return a clean error instead. No behavior change for valid announces.

Bug 2 — Connection.Send silently swallows write errors on a dead conn

connection.go: on "use of closed connection" / "broken pipe" write errors, the loop breaked and Send returned nil → silent message loss. Now returns a new ErrConnectionClosed sentinel (wrapping the underlying error) so callers can detect it and reconnect/retry. Also early-returns on a known-closed connection instead of writing anyway.

⚠️ Behavior change — Send now returns an error on dead-conn writes

Previously Send returned nil in these cases. All four in-repo callers already check the error and handle it well, so this is strictly an improvement:

• requester.go — retries (was: thought it sent, then hung until response-timeout) ✅
• service.go — closes the dead client ✅
• client.go / action_options.go — propagate / log ✅

Callers wanting to special-case it can use errors.Is(err, scamp.ErrConnectionClosed).

Verification

go build / go vet / go test all pass; golangci-lint (v2.12.2/go1.26) 0 issues.

🤖 Generated with Claude Code

[schellj]

Silent message loss is pretty bad.

Both fixes look good.