emileriksenkeev makes long-running legal AI chats far less fragile

- Raise Claude max output tokens 16384 -> 32000 (env CLAUDE_MAX_OUTPUT_TOKENS)
- On stop_reason=max_tokens, continue the loop instead of breaking: run any
  completed tool calls, or push the partial text with a 'continue where you
  left off' nudge. Previously the turn silently ended mid-sentence.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Long tool-JSON generation (wiki page writes, tabular cell generation) streams
no visible bytes for minutes; Cloudflare drops proxied connections after
~100s of silence, surfacing as 'Sorry, something went wrong' while the
backend kept working. Send an SSE comment every 15s; the client parser only
reads 'data:' lines so comments are ignored.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Cache breakpoints on the system prompt and the last message of each request:
each tool-loop iteration now reuses the growing conversation prefix instead of
re-processing all previously read documents (~100k+ tokens per round on wiki
builds). Large latency win between steps and ~10x input-cost cut on long runs.
Request-only copies - stored messages never mutated. CLAUDE_PROMPT_CACHE=0
disables.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

attachActiveVersionPaths / attachLatestVersionNumbers passed every id in a
single PostgREST .in() - beyond ~300 ids the GET URL exceeds the server
limit, supabase-js surfaces the failure only in the discarded error field,
and every document silently loses its storage path. Net effect: any project
with enough documents (RX Labs, 525) got an EMPTY doc context - project
chat saw no documents and wiki builds produced nothing. Batch the lookups
100 ids at a time and log per-batch errors.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Three defects surfaced by a 13-minute, 31-document project-chat review
that ended in a bare "Sorry, something went wrong" while the answer sat
finished in the database.

1. No user message was ever saved. Both chat routes inserted a
   `workflow` field into chat_messages, which has no such column;
   PostgREST rejected every insert and the unchecked result hid it. Drop
   the field and check the error. Reloaded chats now show the questions,
   not just the answers.

2. Cloudflare caps a proxied request at ~10 minutes no matter how often
   the SSE heartbeat fires, so a long agentic run loses its client while
   the backend keeps going and persists the turn. The client now says so
   and polls the saved chat until the answer lands, instead of leaving a
   dead error in place.

3. The client ignored the SSE `error` event entirely, so a server-side
   failure ended the stream as a blank assistant turn. Surface it (and
   send the real message rather than "Stream error").

Also add start/disconnect/finish logging to the project chat route,
which had none - the reason this run left almost no trace.