Extract harness-independent hook protocol

[TensorTemplar]

Closes #50
Addresses the ingestion question in #46

What this does

Implements the stable, abstract hook protocol requested in #50 — independent of Claude Code and OpenCode — so third-party agent tools can feed session events into slopometry without emulating Claude Code hook payloads or patching core.

Protocol design

Following the envelope proposed by @Necmttn in #46, with the closed-source caveat resolved:

• core/protocol/schema.py — EventEnvelope (v1): source, schema_version (validated, fails fast on unsupported), session_id, event_id (required, so backfills are always idempotent), parent_session_id, seq, occurred_at, kind, tool_name, duration_ms, exit_code, error_message, raw.
• core/protocol/kinds.py: canonical EventKind taxonomy (closed, validated: tool_call, tool_result, notification, stop, subagent_stop, subagent_start, todo_updated, message_updated) + open KnownSource constants. The source dimension is an open string — third parties pass any source name without touching core. Kinds are added only when a real adapter needs them (extension policy documented on the enum).
• core/protocol/adapters/: harness-name → canonical-kind maps for Claude (PreToolUse → tool_call, …) and OpenCode (pre_tool_use → tool_call, …). Harness names never reach storage or downstream consumers.
• core/protocol/ingest.py + slopometry ingest --source <name> [--file] [--working-directory]: JSONL envelopes from stdin/file, atomic validate-then-insert, loud failure (exit 2) on invalid lines or source/--source mismatch. Dedupe on (source, event_id) via unique partial index — re-ingesting a trace is a no-op. Sequence numbers continue after the highest stored/batched seq when envelopes omit seq.
• Shared capture_event factory (core/event_capture.py): owns event assembly (sequencing, git-state rule, project detection, tool classification, persistence) for the Claude, OpenCode, and envelope paths — no more near-duplicated construction blocks in the two handlers.

Storage changes

• HookEventType → EventKind everywhere (canonical snake_case values); EventSource closed enum → open str source; HookEvent.event_id added.
• Migration 019: remaps legacy PascalCase wire names and the started/completed dialect from a parallel implementation onto canonical kinds, backfills NULL sources, adds event_id + unique partial index. Numbered 019 because a parallel tree already claimed 016–018 on live databases.

Testing

• New: test_protocol_schema.py, test_ingest.py (envelope validation, idempotency, seq continuation, CLI incl. failure paths), migration-019 test incl. index integrity.
• Full suite: 855 passed standalone on this branch (ruff clean, mypy no new errors).
• Verified live: installed per README, migration 019 applied cleanly to a 910k-event production DB, ingest + re-ingest + solo ls/solo show all correct on migrated data.

For reviewers from #46

@botbotfromuk — this answers your question: ingest is an explicit import call (slopometry ingest), not a watched path. A collector like your SlopometryCollector can now write envelope JSONL and pipe it in; source + event_id makes retroactive convert_trace_to_slopometry() backfills safe to re-run. Mapping tick_start/decision/action → notification is the pragmatic choice; keep finer distinctions in raw until real consumers need first-class kinds (see the extension policy in kinds.py).

@Necmttn — the closed AbstractEventSource caveat from your review is resolved: source is an open validated string, KnownSource is constants-only, nothing gates on membership.

[TensorTemplar]

Closing: superseded by a delta approach. Since this branched, #54 landed the abstract protocol on main (AbstractHookEvent, AbstractEventType, adapters, dispatch, emit-event) — this PR unknowingly re-implemented it with a divergent taxonomy and is CONFLICTING as a result. The genuinely additive parts (open source dimension, stable external envelope with required event_id, idempotent JSONL backfill) will be re-submitted as a small PR on top of current main instead. Apologies for the churn — and to reviewers from #46: the ingest contract question will be answered by the follow-up PR.