[Dimoshka]

Refs #33

PR #34 merged before this review round finished, so the six review-fix commits
that were still being written sit on top of the merged head (2742661) and land
here instead. They are test-strength fixes and small residuals, not a rewrite:
no behaviour changes beyond the empty-reply draft fix.

Commits

• 08f1dd2 fix(app): clear the draft when a terminal reply is empty — a terminal event with an empty reply now clears the in-flight draft instead of leaving stale text in output, matching the state contract.
• 146b8f4 test(app): exercise the fold guard and empty-reply paths — the "folds once" test now actually delivers a duplicate terminal to the bloc, and the zero-length vs whitespace-only reply paths are pinned separately.
• e35afab test(app): cover follow-along scrolling in the conversation view — follow-along is now covered both ways: a token follows when the reader is at the bottom, and does not when the reader has scrolled up.
• fd4efbc test(core): pin the chat role wire values — the ChatMessage role constants and knownRoles are asserted as exact strings, so the wire protocol cannot drift silently.
• ddcc418 refactor(cli): use the ChatMessage role constants — the CLI build of the message list now uses the shared role constants instead of duplicated literals.
• 2a34a80 docs: refresh the measured suite counts — documented counts refreshed to the measured 1199.
• 48b33b4 docs: record PR feat(chat): a real conversation transcript in the Chat tab #34 as merged and test(app): tighten the chat transcript coverage after review #35 as the follow-up — docs/status.md, AGENT.md and AGENTS.md no longer describe the transcript as unmerged, and the Windows aa35432 figures are labelled as a historical measurement.

Test plan

• dart run melos run check — green: gen, catalog check, analyze, tests, format, CLI build.
• dart run melos run gen:catalog:check — green; catalog up to date (362 parameters, digest sha256:1838d6c8…bdadb8).
• cd apps/llama_pulse && flutter build linux --debug — green; build/linux/x64/debug/bundle/llama_pulse produced.
• Measured suites: core 800 (+1 environment-gated skip), CLI 233, app 166 — 1199 total.

Note on coverage strength: the fold-once test fails if the draft folding in
_commitDraft is removed (verified by temporarily removing it), so it guards
real behaviour. The _terminalSeen check in the event gate is redundant with
the request-id gate — after a terminal the request id is null, so the duplicate
is already ignored — and the test passes with that check removed as well
(verified).