fix(gateway): restore session transcript fallback

[LeonSGP43]

Summary

• restore JSONL transcript writes in SessionStore.append_to_transcript and rewrites in rewrite_transcript
• fall back to the JSONL backup when DB transcript rows are missing or unreadable
• add regression tests for DB-empty fallback and skip_db transcript persistence

Closes #33563.

Testing

• uv run --frozen pytest -o addopts= tests/gateway/test_session.py -k "LoadTranscriptFallback or RewriteTranscript"
• uv run --frozen pytest -o addopts= tests/run_agent/test_860_dedup.py -k "skip_db or default_writes_to_sqlite"
• uv run --frozen ruff check gateway/session.py tests/gateway/test_session.py tests/run_agent/test_860_dedup.py
• git diff --check

Proof

• .paperclip_artifacts/quality-gate/proof-71574eb3ceb2ffb505bb8f98ef96951c14904784.json

[teknium1]

Thanks for taking a crack at this, and the regression tests are clean work. Unfortunately we can't take this one, and the reason is architectural rather than a quality issue.

This PR restores JSONL transcript writes in append_to_transcript/rewrite_transcript and re-adds a JSONL fallback in load_transcript. But the JSONL fallback was deliberately removed in commit 024a8e3ee ("refactor(gateway): drop JSONL fallback in load_transcript") — state.db is the canonical session store by design, and pre-DB sessions were already migrated into it. Re-adding the JSONL path reintroduces dual-write/dual-read divergence (the #860 duplicate-write class) and revives an approach we intentionally moved past.

The underlying issue this targets (#33563) was a stale premise plus a real-but-already-fixed corruption bug:

• The "missing JSONL" the reporter saw is expected — persistence flows through state.db per turn (finalize_turn → _persist_session → _flush_messages_to_session_db), not JSONL.
• The actual data loss was state.db FTS corruption silently swallowing writes — now detected and auto-repaired on open via fix(state.db): recover from malformed sqlite_master so hidden sessions reappear #43149 (malformed sqlite_master) and fix(state): detect and repair FTS write corruption that silently drops gateway history (#50502) #52798 (FTS write-corruption, the exact class in that report). Both landed after Gateway Telegram sessions lose context after idle TTL eviction — session JSONL never written, state.db corruption swallows writes silently #33563 was filed.

So the right fix for the data-loss symptom is the corruption detection/repair that's already on main, not restoring a parallel JSONL store. Closing #33563 as resolved alongside this.

Appreciate the contribution — the DB-empty-fallback test instinct was sound; it just points at a layer we've intentionally collapsed onto state.db.