fix(stream): preserve per-segment ids so assistant parts keep chronological order

[vutuanlinh2k2]

Problem

normalizePersistedPart strips id from text/reasoning parts, so getPartKey collapses every segment to text:current / reasoning:current: all reasoning merges into one part, all text into one part, and interleaving (reasoning → tool → reasoning → text) is destroyed at persist time. finalizeAssistantParts then stamps the full final text into every text part.

Harness adapters already emit stable per-segment part ids (OpenCode native ids; claude-code streaming path assigns a UUID per thinking block) — the fidelity is thrown away here.

Change

• normalizePersistedPart: text/reasoning parts carry id (or partId) through when present. Absent id = legacy single-segment flow, unchanged — an id is never invented.
• mergePersistedPart: an empty text snapshot no longer erases accumulated text (now matches the reasoning branch).
• finalizeAssistantParts gains a per-id contract with the invariant concat(text segments) === final text:
  • no text parts → append finalText as one part (unchanged)
  • no text part has an id (legacy) → overwrite with finalText (byte-identical to today)
  • per-id segments: equal → untouched; finalText extends the stream (e.g. an appended failure diagnostic) → id-less remainder segment appended; diverged → collapse to one authoritative segment at the last text position (finalText is what content/post-processing consume — it must win, while reasoning/tool chronology survives)

Back-compat

Id-less flows (creative-agent call sites, old persisted rows, router-style text streams) behave exactly as before — covered by dedicated tests.

Tests

New tests/stream-normalizer.test.ts: id preservation/no-invention, per-id keying vs *:current fallback, per-id merge semantics, all finalize branches, and end-to-end order integrity for reasoning → tool → reasoning → text.

Context

Foundation for tangle-network/gtm-agent#448 (chronological chat transcript); gtm consumes this via the auto-published patch bump. Note: will conflict textually with fix/dangling-tool-terminalization if that lands too — whichever merges second needs a trivial rebase (the terminalize wrapping is tool-only, orthogonal to these branches).