Expand the anonymous product event catalog

[TheGreatAxios]

Replaces #416, rebased onto cl-5743-batch-the-telemetry-transport (session id + batched transport).

The defect this fixes

#416 sent five user-authored identifiers to PostHog, each behind a comment asserting it was already a safe enum. The comments were false: permission_kind was the raw tool id, which for MCP tools is mcp__<server>__<tool> where <server> is the operator's own settings key; skill names, plugin ids, and agent profile ids are all project- or author-chosen. A sixth of the same class was found and fixed: slash_command's command_name came from a registry plugins also write into.

Every such value is now mapped to a fixed first-party enum at the emission site (src/telemetry/classify.ts) and reported as custom when it matches nothing, with mcp as its own bucket for permission_kind. There is no first-party list of skills or plugins to match against, so skill_used carries no name at all and plugin_loaded carries only origin — the discovery tier.

Also

• Consumes getSessionId() from the base branch instead of re-declaring it.
• parent_session_id dropped: sub-agents run in-process on the same session id, so it always restated session_id.
• plugin_used dropped: it fired at startup over every enabled plugin and observed nothing about use, which plugin_loaded already measures honestly.
• compaction no longer fires on no-ops, and its inverted trigger label is replaced by the compactor's actual mode.
• crash reports only language-defined error types by name; an application or plugin error subclass buckets to custom.
• The two byte-identical permission_prompt sites in the gate are one helper. They are mutually exclusive and buildRequests yields at most one request per call, so a prompt is counted once.
• Emission takes Telemetry as an injected dependency, matching Emit PostHog AI observability events in privacy mode #417, rather than reading the process-wide handle.

Test plan

tests/unit/telemetry-product-events.test.ts drives each real emission site with a deliberately identifying input (mcp__acme-internal__deploy, a skill directory named after a company, a plugin manifest id containing a path, a project agent profile, a plugin-registered command) and asserts both the bucketed value and that the raw string appears nowhere in the serialized request body.

• bun run typecheck — 0 errors (base: 0)
• bun run build
• bun run test — 4278 pass, 1 unrelated flake in src/tui-opentui/markdown-rows.test.ts (passes in isolation, no file it touches is in this diff)
• bun test ./src ./tests ./evals --randomize — 4279 pass, 0 fail

---

Branch updated by rebase (authorized)

This PR branch was rebased onto main and force-updated. The pre-rebase head was 56ae4c98a. Those commits remain fetchable from that SHA.

This updates the branch behind this pull request; it does not bypass review. The PR still merges through review as normal. A plain push is refused after a rebase rewrites history, which is the only reason a force update was needed. Authorized by the operator.

Why the rebase was necessary

#412 was squash-merged to main, and #423 was squash-merged into its own base branch cl-5727-posthog-identity-session-linking after that branch had already merged. The #423 squash therefore never reached main: main carries three telemetry events, no /batch/ endpoint and no discard().

Rebasing this branch with cl-5743-batch-the-telemetry-transport as the upstream would have silently dropped the entire batching transport and the opt-out discard, without raising a conflict. The rebase instead used the top of the squashed #412 commits as the upstream, so only genuinely duplicated work was dropped. The batching commits replayed with zero conflicts, which is the evidence that this was the correct shape.

This PR is currently the only path by which the batching transport and the opt-out discard reach main. cl-5727-posthog-identity-session-linking must not be deleted until this lands.

[linear-code]

CL-5722