Compress session_events tool_input + usage (v2 of #43 — the real ~380MB bulk)

Implemented Brief #51 on branch flower/compress-tool-input. Commits: - a83eebb Compress session event JSON payloads - 3d91a51 Backfill compressed session event payloads Scope: - Added session_events.tool_input_gz + usage_gz migration (MySQL LONGBLOB, sqlite binary for tests) while keeping legacy JSON columns. - Extended SessionEvent compression accessors/mutators to tool_input + usage, sharing the #43 config gate and covering relationship creates, update(), bulk insert(), and partial selects. - Updated SessionProjectTouchService's tool_input presence query to include tool_input_gz and verified no other raw app-side session_events tool_input/usage presence checks needed adjustment. - Extended flower:compress-session-events to backfill text, tool_input, and usage idempotently, and added --optimize for MySQL OPTIMIZE TABLE after forced backfill. - Added sqlite coverage for JSON round-trip, legacy fallback, config-off writes, bulk ingest + SegmentSession reads, compressed tool_input attribution, and command dry-run/force/optimize behavior. Verification: - ~/bin/php vendor/bin/pint app/Models/SessionEvent.php app/Models/Builders/SessionEventBuilder.php app/Services/Attribution/SessionProjectTouchService.php app/Console/Commands/CompressSessionEvents.php database/migrations/2026_07_01_190000_add_compressed_json_to_session_events_table.php tests/Feature/SessionEventCompressionTest.php tests/Feature/CompressSessionEventsCommandTest.php tests/Feature/FoundationTest.php - ~/bin/php artisan test --filter='SessionEventCompressionTest|CompressSessionEventsCommandTest|TranscriptBuilderTest|TranscriptIngestTest|SegmentSessionTest|FoundationTest|UmbrellaAttributionTest' - ~/bin/php artisan test Safety: did not run migrate, tinker, or flower:compress-session-events against the worktree/live DB. Horizon reload flag: this extends IngestSession/SegmentSession event storage/read paths, so MAIN should graceful reload Horizon after merge.

Started implementation on branch flower/compress-tool-input from master. Grounded with recall_brief #51 and prior #43 recall. Safety boundary acknowledged: sqlite php artisan test only; will not run migrate, backfill, flower:compress-session-events, or mutating tinker against the worktree/live DB. Will flag Horizon graceful reload because this extends IngestSession/SegmentSession event storage/read paths.

## Goal v1 (#43) compressed `session_events.text` → 71 MB `text_gz`. But the REAL bulk is **`tool_input` = 359 MB** (raw tool-call args — file contents, bash output) + **`usage` = 21 MB**, both uncompressed. Compress them the same way. Measured on MAIN: session_events .ibd ~700 MB; after this + OPTIMIZE it should drop to ~200 MB. ## Approach — mirror #43 v1 exactly 1. **Schema:** migration adds nullable `tool_input_gz LONGBLOB` + `usage_gz LONGBLOB` (keep the legacy `tool_input`/`usage` JSON columns for fallback). 2. **Transparent accessor/mutator on `App\Models\SessionEvent`** (same shape as the existing text_gz one): SET gzcompress()es the JSON string into the _gz column (legacy col null); GET gunzips + json-decodes (fallback to the legacy column when _gz is null). Cover the bulk `insert()` path (IngestSession) like v1 did. Same `config('flower.ingest.compress_events')` gate. 3. **Fix the one SQL presence-check:** `app/Services/Attribution/SessionProjectTouchService.php:~206` does `->whereNotNull('tool_input')` — after compression that's null, so update it to also/instead check `tool_input_gz`. Grep for any other raw `tool_input`/`usage` SQL and adjust. 4. **Extend `flower:compress-session-events`** to also backfill tool_input→tool_input_gz and usage→usage_gz (idempotent; it already prints before/after size). Keep dry-run default + --force + --batch. 5. **Reclaim:** add an `--optimize` flag (or document) that runs `OPTIMIZE TABLE session_events` after backfill — InnoDB doesn't return UPDATE-freed space automatically, so the .ibd only shrinks after a rebuild. (Orchestrator will run backfill + OPTIMIZE on MAIN.) ## Acceptance - tool_input_gz + usage_gz LONGBLOB migration; transparent accessors (compress on write, gunzip+json-decode on read, legacy fallback, bulk-insert covered). - SessionProjectTouchService presence-check updated; no other raw tool_input/usage query broken. - Backfill command compresses all three columns; idempotent; before/after size report; --optimize flag. - Tests (sqlite): round-trip for tool_input+usage (array in → compressed → array out), legacy fallback, bulk-insert, the touch-service query still works post-compression. Suite green; pint. - **FLAG on merge:** touches SegmentSession/IngestSession (Horizon jobs) → graceful Horizon reload. `Brief:` trailer with THIS brief's id. ## Guardrails ⚠️ Worktree .env may point at MAIN + this touches session_events — **sqlite tests ONLY**; NO backfill/migrate/mutating tinker against the live DB. The orchestrator runs the backfill + OPTIMIZE on MAIN. ## Provenance v2 of #43 (text v1 shipped). Measured on MAIN 2026-07-01: tool_input=359MB (actual bulk), usage=21MB, text_gz=71MB, .ibd ~700MB. Operator: "yes dispatch v2".