feat: skip autoRetry for sendMessageDraft to stop 429 amplification

[fitz123]

Summary

Closes #117. Wrap autoRetry in a method-gated transformer that bypasses retries for sendMessageDraft only. Drafts are cosmetic fire-and-forget calls — a 3-10 s late retry lands on a stream that has already moved on, and the 5× amplification turns one rate-limited draft into five log lines and five Prometheus error-counter increments. Every other Telegram API method retains the full AUTO_RETRY_OPTIONS retry behavior unchanged.

Also fixes the misleading "Drafts are cosmetic — no rate limits" comment in bot/src/stream-relay.ts that originally motivated the 300 ms debounce, and adds a one-line README caveat that bot_telegram_api_calls_total{method="sendMessageDraft"} no longer counts per-attempt (one call = one increment, vs other methods' up-to-5).

Plan and pipeline

• Plan: docs/plans/2026-05-26-telegram-draft-autoretry-skip.md
• Ralphex pipeline ran: implementation + 7-agent multi-review (claude) passed clean, then a 2nd-pass claude review picked up two findings — one false positive (rejected), one acted on (README per-attempt caveat). Codex external review: NO ISSUES FOUND.
• Pipeline terminated on the final "claude evaluating codex findings" step due to a Claude session-window limit; all substantive work was complete by that point.

What's NOT in scope

• Changing DRAFT_DEBOUNCE_MS (Option A from sendMessageDraft 429 cascade during long streaming replies in DMs #117). Deferred until residual sendMessageDraft 429 ratio is measured against the new bot_telegram_api_calls_total counter (deployed in feat: chat_id in API error logs + bot_telegram_api_calls_total counter #118).

Test plan

• cd bot && npx tsc --noEmit — passes
• cd bot && npm test — 1068 / 1068 pass
• After merge + restart: trigger a long streaming reply in a DM, confirm Rate limited: method=sendMessageDraft warn lines drop ~5× and sendMessage/sendChatAction still retry on 429
• After merge + restart: curl -s http://127.0.0.1:9091/metrics | grep bot_telegram_api_calls_total — verify sendMessageDraft rows still increment (one per logical draft, not per retry)

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR reduces Telegram 429 log/metric amplification during long DM streaming replies by disabling @grammyjs/auto-retry specifically for the cosmetic sendMessageDraft API method, while keeping existing retry behavior unchanged for all other Telegram API methods.

Changes:

• Add a method-gated API transformer that bypasses autoRetry for sendMessageDraft only.
• Update inline docs/comments and README to clarify draft rate-limits and the resulting metric semantics.
• Add unit tests that assert sendMessageDraft is not retried while sendMessage and sendChatAction still retry on 429.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
README.md	Documents the sendMessageDraft exception to per-attempt counter behavior due to retry bypass.
docs/plans/2026-05-26-telegram-draft-autoretry-skip.md	Adds an implementation plan and rationale for skipping autoRetry on drafts.
bot/src/telegram-bot.ts	Introduces createDraftSkipAutoRetryTransformer and wires it into the API transformer chain.
bot/src/stream-relay.ts	Corrects the draft debounce comment to acknowledge per-chat rate limits and the retry bypass behavior.
bot/src/tests/telegram-bot.test.ts	Adds tests verifying draft bypass and continued retries for other methods.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.