feat: fold quoted/replied-to message into the run prompt (Telegram + VK)

[zaytcevcom]

Summary

When a user replies to or quotes an earlier message, the Telegram and VK adapters previously forwarded only the user's new text to the agent — the referenced original was lost. A reply like "Про это" ("about this") therefore reached the run with no idea what "this" was.

This folds the quoted original into the run prompt on both transports, through a single shared helper, across every input path (text, voice and media). The quoted text is framed as reference context — data, not instructions — and placed before the user's own message, which stays the operative instruction. A non-reply message is byte-for-byte unchanged.

Changes per acceptance criterion

• AC1 — shared helper (core/chat/media.go): new QuotedPrompt(quotedAuthor, quotedText, userText), mirroring the existing DocumentPrompt/PhotoPrompt pattern. It places the quoted block first (framed as reference data), the user's message last; adds an author clause only when known; returns userText unchanged when the quote is blank (the no-op that keeps non-reply traffic identical); and truncates an over-long quote rune-safely at maxQuotedRunes = 2000 while never touching userText.
• AC2 — Telegram (cmd/flock-telegram/main.go): quotedContext extracts the quoted original — the exact highlighted portion (msg.Quote.Text) if present, else the replied-to message's text, else its caption, else a generic [media] reference — and folds it via QuotedPrompt into the text, voice and media paths. Only the bot's own message (matched by id) is labeled "the assistant"; any other sender is attributed by name.
• AC3 — VK (adapters/vk/longpoll.go, adapters/vk/receiver.go): the message object now parses reply_message and fwd_messages; quotedContext folds the replied-to message (preferred) or the first forwarded message — one level only — via the same QuotedPrompt, with a coarse assistant-vs-participant author label derived from from_id's sign (no extra users.get call).
• AC4 — edges: a media-only quoted message yields a non-empty [media] reference (never a silent drop); an empty user message is still dropped (a bare reply with no new words does not run); a very long quote is truncated. Non-reply traffic and all pre-existing tests are untouched.

How it was verified

Run through the repo's own gate inside the dev-tools image (Go 1.26.5 / golangci-lint 2.11.4) — exactly what CI runs:

• task tests — go test -race ./..., all packages green, including the pre-existing suites (regression guard) and new unit tests: QuotedPrompt (both texts distinguishable, blank → no-op, rune-safe truncation with a UTF-8 validity check); the Telegram handler (quote-wins-over-reply, own-bot vs third-party-bot label, media-only reply, non-reply unchanged, quote+caption+photo carries the quote and keeps the vision image); the VK receiver (reply, forward first-only, media-only reply/forward, non-reply unchanged).
• task lint — 0 issues. task build — green.
• Not verified: live end-to-end behaviour against real Telegram/VK Long Poll updates (see residual risks).

Pre-PR review

One adversarial review pass on the diff → clean round (no blocker/major), risk: low. It confirmed the load-bearing points: the no-op is byte-for-byte on non-reply, truncation is genuinely rune-safe, there is no double-wrap (media/voice paths return before the text-path wrap; each path calls the helper exactly once), VK pointers/slices are nil-guarded and read one level only, and prompt-injection framing is present.

Minor findings raised and fixed in this branch before opening:

• The author label originally treated any bot as "the assistant" (from.IsBot), which would attribute a third-party bot's words to ourselves. Narrowed to the bot's own id (the same check as replyToBot); added a test for the third-party-bot case.
• Added a test for the highest-preference msg.Quote branch (Telegram's native quote), previously unproven.
• Broadened the media placeholder to any text-less reply (covers polls/locations, not just enumerated media types); dropped the now-redundant replyHasMedia enumeration.
• Rune-count before allocating on the truncation path.

Residual risks / worth a human eye:

• Live behaviour not exercised. Verified by unit + -race suite, not against a live bot. Worth a smoke test on each transport: a plain reply, a native quote selection, a forward, a reply to a photo/voice, and a reply to the bot's own message.
• Quoted media is not described. Replying to a photo/file yields the literal [media] marker — not a description or the image itself (no vision on the quoted target). Intentional for v1.
• VK author label is coarse ("a participant", no name), and VK JSON shapes are assumed from the API (validated by struct + crafted-JSON tests, not a live Long Poll payload). If VK nests differently, parsing degrades to a safe no-op (the feature simply would not fire).
• Prompt-injection residue. The quoted block is framed as data, but the model still sees text authored by another (allow-listed) user; the framing mitigates, it does not eliminate.
• The PR-time multi-arch image builds (Telegram/VK Dockerfiles) run fresh in CI; no Dockerfile was touched here.