XERK-516 [Qwen K]: session migration + resume for qwen

[xerhab]

What & why

Extends XERK-101 session migration + resume to the qwen (Qwen Code) runtime — the dsh [K] (XERK-475) analogue. Parent: XERK-504.

The load-bearing difference from dsh: qwen is Claude-shaped, so its native log (~/.qwen/projects/<slug>/chats/<id>.jsonl) IS what qwen --resume <id> reloads from — there is no separate durable store like dsh's DSH_SESSIONS_ROOT. So a qwen migration carries that native log itself, not a distinct store.

Resume (process-death --resume, boot-adopt) was already wired by [Qwen B]/[Qwen C] (_launch_qwen(resume=True) + tail at native-log EOF); this ticket pins it. The net-new work is migration.

Changes (all in agent/hub-agent.py)

• _qwen_native_log(sid) — glob-locate the native log (slug-rule-independent, like _qwen_runtime_file/the tail).
• _qwen_store_dest(cwd, sid) — target native-log path under the target cwd's slug. Verified: qwen's cwd→slug rule is _project_slug (every non-alnum→-), not the G0 doc's imprecise /→-.
• export_session — for a qwen session, packs the native log under the reserved .qwen-store/chat.jsonl prefix, truncated to its last complete line (a live log is appended by its process). Un-droppable resumable data — export refuses cleanly over MIGRATION_BLOB_MAX rather than shipping un-resumable.
• import_session / _unpack_transcript — routes the .qwen-store/* member to the single target file (never a dir — the shared chats/ dir holds other sessions' logs), with the same path-escape hardening as the dsh store. Dropped on a claude fallback (no qwen on target, via the existing agent_type_configured guard).
• _reconcile_qwen_store_cwd — re-points the cwd carried on every native-log row (qwen keys on cwd — upstream issue #2373) from the source to the localized worktree; no-op on same-mount; never raises.
• _launch_qwen — re-validates the carried model against the target host's config (kept only if QWEN_IDENT_RE-valid, else host default QWEN_MODEL); endpoint/key are always the target's.

The <slug>/<sid>/qwen/ raw-archive mirror is the display/metrics feed and is never carried by migration (the target rebuilds it) — the [K] correction the ticket names.

Verification

• Unit tests: 2123 agent tests pass. Added 6 qwen cases to TestMigrateSession (pack/unpack round-trip + tail truncation, stray-drop on claude import, cross-mount place+rekey end-to-end, claude-fallback drop, slug-rule pin, same-mount no-op).
• QA (qa agent): PARTIAL, zero defects, zero security findings. Everything runnable was driven:
  • Slug rule confirmed 12/12 against real on-disk ~/.qwen/projects/ logs (incl. the git/.turma→git--turma case the G0 rule would break).
  • _unpack_transcript path-escape hardening confirmed with hostile tar members (traversal/absolute → ValueError; symlink/dir → skipped; member relpath discarded so no escape).
  • _reconcile_qwen_store_cwd confirmed across 11 edge cases: rows & fields preserved, unparseable lines kept verbatim, zero raises, correct same-mount no-op.
  • Export blob-ceiling refusal confirmed; mutation testing confirmed each test catches its regression.
  • Host-proof gap (the PARTIAL, exactly as the ticket states): a real cross-host move actually resumed by qwen --resume over the placed+rekeyed log is not runnable in one worktree and qwen isn't in CI — the same footing [Qwen C]/[Qwen E] shipped on. Low-risk residual: reconcile reserializes matching rows (valid JSON, same values), unproven only if qwen does byte-level per-row integrity checks; dsh's analogous header rewrite works against real dsh.

Docs

• [Qwen K] section added to .claude/rules/qwen.md (32,971 chars, under the 40k cap).
• Migration note added to .claude/rules/agent-archive.md.