fix(relay): stop panic when reacting to a project root or comment

[BradGroux]

What users saw

Tapping a reaction on an issue, pull request, or a comment on one could panic the relay's ingest worker. The reaction row was inserted before the panic, so the original publisher could see a failed request even though the event had already been persisted.

Why it happened

Reactions (NIP-25, kind 7) derive their channel scope from the target event. Project events such as issues and pull requests do not carry an h tag, so derive_reaction_channel returns NoChannel and channel_id is None.

The conformance-trace emission on the reaction path assumed channel_id was always present:

channel: channel_label(channel_id.expect("reaction path has channel")),

That assumption predates reactions on project events. Because persistence completes before the trace action is built, the panic happens after the database write.

What changed

Extracted the three-way (channel_id, was_inserted) match already used by the general message-write path into a shared write_trace_action helper:

• (Some(ch), true) becomes WriteInsert
• (Some(ch), false) becomes WriteDuplicate
• (None, _) becomes WriteInsertGlobal

Both the reaction path and the message-write path now call this helper. Channel-less project reactions use the existing global-write trace vocabulary instead of unwrapping a missing channel.

How this was tested

Added four unit tests covering the complete helper matrix:

• channel-less insert emits WriteInsertGlobal
• channel-less duplicate emits WriteInsertGlobal
• channel-bearing insert emits WriteInsert
• channel-bearing duplicate emits WriteDuplicate

The channel-less insert is the exact affected path. A normal retry of the same active reaction exits earlier through ReactionEventInsertOutcome::Duplicate; the channel-less duplicate case is defensive coverage for the helper contract rather than a claim about that retry path.

Validation at f611f96e7:

./bin/cargo fmt --all -- --check
cargo test -p buzz-relay --lib handlers::ingest::tests
cargo clippy -p buzz-relay --all-targets -- -D warnings

All 164 ingest tests passed, and formatting and strict Clippy checks completed successfully.

Scope and non-goals

• Does not change reaction persistence or channel derivation.
• Does not change duplicate-reaction behavior.
• Only replaces the unsafe trace-action construction and shares the mapping with the message path.

Closes #4936.

[BradGroux]

Review update: I traced the insert and duplicate control flow, corrected the PR description, and pushed the required formatting cleanup. The functional fix remains narrowly scoped to trace-action construction. A normal retry of the same active reaction exits earlier through ReactionEventInsertOutcome::Duplicate; it does not re-enter the trace helper, so the previous claim that every retry panicked was inaccurate.

At f611f96e7, rustfmt, all 164 relay ingest tests, and strict Clippy pass. I also started the repository-wide test recipe: its non-integration suites advanced successfully, but the integration-service gate could not start because no Docker daemon was available in the review environment.

[BradGroux]

Rebase and accuracy review (2026-08-07)

Rebased onto current main (2b873cf20). Clean fast-forward — 52 commits applied, no conflicts.

Branch state: 639c8c256 (was f611f96e7) Mergeable: ✅ MERGEABLE · BLOCKED (branch protection — awaiting maintainer review)

Accuracy review

The fix addresses a real panic: reactions (kind 7) on project events (issues, PRs, comments) derive NoChannel because project events lack an h tag, and the conformance-trace emission called channel_id.expect("reaction path has channel") after the database write had already completed.

The resolution extracts the three-way (channel_id, was_inserted) match into a shared write_trace_action helper, used by both the reaction path and the message-write path. Channel-less reactions now use WriteInsertGlobal instead of unwrapping. Four unit tests cover the complete matrix.

No unwrap()/expect() remains on the reaction trace path. Scope is limited to ingest.rs. No new dependencies.

Closes #4936.