fix: recover from a Letta conversation the backend no longer has

[EthanThatOneKid]

Problem

A session's stored conversation id is replayed on every later turn. When the local
Letta backend no longer has that conversation, every request for that session fails
permanently — no retry ever succeeds, because the same dead id is re-sent.

Data's Discord session held a pre-cutover con_ id that the local backend never had
(it names conversations local-conv-N). Every Discord mention returned 502:

letta -> exit 1: Conversation con_QPvFAwS14xUBFIIb not found

Observed live: a real mention (1553268586628255786) failed to answer at
2026-09-26T04:54:50Z for this reason.

Fix

• Detect the backend's missing-conversation failure (lib/conversation-recovery.ts),
  which prints Conversation <id> not found on stderr and exits 1 with no stdout.
• In askLetta, drop the stale mapping and retry once against a fresh conversation,
  so a stale id costs one extra turn instead of bricking the session.
• The detector is only consulted when no result object was parsed from the turn, so
  prose in a real answer cannot trigger a spurious retry.

Verification

• npm run typecheck clean; npm test 26/26 pass (2 new tests).

• Isolated end-to-end run of channels/http/index.ts against the real local Letta
  agent, seeded with the stale id as a session mapping:

  request	POST /ask {session:"stale-test"}
  log	conversation con_QPvFAwS14xUBFIIb is gone; starting a fresh one for session stale-test
  response	200 with a fresh local-conv-63
  state	mapping rewritten to the new id

• The live host's stale entry was cleared and data-http restarted; a /ask on the
  real Discord session (discord:1531147245318176848) now returns 200 with a fresh
  conversation, where it previously 502'd.

[wazoocomputer]

This pull request addresses an issue where a session's stored conversation ID, when no longer recognized by the Letta backend, would cause all subsequent requests for that session to fail. The fix introduces a mechanism to detect this "missing conversation" failure, drop the stale conversation ID, and retry the request with a fresh conversation.

File	Change	Summary
channels/http/index.ts	modified	Implements conversation recovery logic and retries requests with fresh conversations when a stale ID is detected.
lib/conversation-recovery.test.ts	added	Adds tests for the isMissingConversation utility function.
lib/conversation-recovery.ts	added	Defines isMissingConversation to detect missing conversation errors from the Letta backend.

Reviewers, please start by examining lib/conversation-recovery.ts and its associated tests, then move to channels/http/index.ts to see how the recovery logic is integrated.

[wazoocomputer]

This pull request addresses an issue where a session's stored conversation ID, when no longer recognized by the Letta backend, would cause all subsequent requests for that session to fail. The fix introduces a mechanism to detect this "missing conversation" failure, drop the stale conversation ID, and retry the request with a fresh conversation.

File	Change	Summary
channels/http/index.ts	modified	Implements conversation recovery logic and retries requests with fresh conversations when a stale ID is detected.
lib/conversation-recovery.test.ts	added	Adds tests for the isMissingConversation utility function.
lib/conversation-recovery.ts	added	Defines isMissingConversation to detect missing conversation errors from the Letta backend.

Reviewers, please start by examining lib/conversation-recovery.ts and its associated tests, then move to channels/http/index.ts to see how the recovery logic is integrated.

[EthanThatOneKid]

Added the re-brick guard as recovered (40bf2d5) and documented the boundary.

Boundary — what this does not cover. Recovery here handles conversation loss. A wiped backend also loses the agent, which surfaces differently — Agent agent-local-... not found instead of Conversation ... not found — and the detector deliberately ignores it. That case still needs letta --backend local agents create plus a DATA_LETTA_AGENT_ID update. Now stated in channels/http/README.md.

Verification. A stub letta reproduces the backend's shape (exit 1, empty stdout, Conversation <id> not found on stderr) and a fresh turn that answers but omits conversation_id from its JSON — the precondition for the re-brick. Seeding a dead id for discord:test and running one /ask:

• against committed ea6ba1d: dead id written back with a fresh timestamp → RE-BRICKED
• against this branch: conversations.json left as {}, conversationId: null → fixed

Typecheck clean, npm test 26/26.