v0.2.0 — Per-engine fan-out, service binding, status column

A correctness + architecture pass driven by an end-to-end debugging session on the production fork. Every fix is described in terms of what the prior version got wrong.

Added

• migrations/0004_response_status.sql: status (ok / failed / skipped) and error_message columns on prompt_responses, plus an index on status. Failed engine calls used to write raw_response='ERROR: ...' rows that downstream scoring treated as real zero-mention hits. They're now explicit status='failed' rows that aggregates exclude.
• POST /admin/run-engine — single-engine handler. /admin/run-live self-fetches into it once per engine so each engine runs in its own worker invocation with its own free-plan 50-subrequest budget. Idempotent on (run_id, engine).
• POST /admin/cleanup-failed-runs — one-shot deletion of legacy polluted rows that were written before the status column existed. Idempotent, optional brand_id filter.
• bulkCacheGet helper in src/db.ts — one D1 read for N prompt hashes instead of N separate reads.
• persistEngineRun helper in src/db.ts — one batched D1 write at the end of an engine run instead of per-prompt inserts + cache puts + per-chunk run updates.
• EnginePromptResult interface — engines collect their per-prompt results in memory and flush in one batch.
• runEngineInProcess in src/engines.ts — the in-process dispatcher used by /admin/run-engine.
• Fetcher service binding (env.SELF) in Env and EnginesEnv, plus "services" block in wrangler.example.jsonc. The per-engine fan-out goes through env.SELF.fetch() because a public-URL fetch back to your own workers.dev hostname trips Cloudflare's "Worker called itself" guard (error 1042) and never lands.
• SELF_URL env var in wrangler.example.jsonc — the cron path needs to know the worker's canonical URL since it has no inbound request to derive an origin from.

Changed

• Each engine client (src/openai.ts, anthropic.ts, perplexity.ts, gemini.ts, ai-overviews.ts) now uses the bulk pattern: hash all prompts up front, bulkCacheGet once, run LLM fetches in CONCURRENCY=5 chunks, collect EnginePromptResult[], flush via persistEngineRun. Per-invocation subrequest count drops from ~89 to ~26.
• getLatestCompletedRun anchors on EXISTS(prompt_responses with status='ok') instead of runs.status='completed'. Partially-finished runs still surface their data instead of silently disappearing from MCP tool output.
• check_visibility, compare_competitors, and get_content_gaps use run.completed_at ?? run.started_at so they accept in-progress runs.
• get_visibility_history and get_citations SQL queries filter pr.status = 'ok' in the join/where clause so failed/skipped rows can't pollute aggregates.
• /admin/run-engine INSERT OR IGNOREs its runs row before calling the engine — D1's cross-region replication may not have caught up with /admin/run-live's INSERT INTO runs by the time the self-fetch lands, and a missing parent row would crash the engine's batched insert with FOREIGN KEY constraint failed.
• runEngines accepts an optional request argument and falls back to new URL(request.url).origin when SELF_URL isn't set, so HTTP-triggered code paths work even without the env var.

Fixed

• Subrequest-cap row loss: pre-fix, every engine deterministically wrote 10 rows out of 20 because chunk 3 hit the 50-subrequest cap. Now writes complete in one batch under the cap.
• Silent row drops from failed Gemini / AI Overviews calls: pre-fix the catch block's insertPromptResponse(failed) was itself a D1 subrequest that could fail, leaving zero rows for the prompt. The bulk batch holds failure results in memory and writes them in the same batch as the successes.
• Visibility tools dropping engines whose runs got stuck at in_progress: relaxed getLatestCompletedRun surfaces them.