feat(session): session server v2 — opt-in trajectory-tree serving; v1 preserved byte-exact

[guapisolo]

Motivation

The session server needs multi-lineage trajectory support (retries, compaction with carried assistants, subagent forests). Earlier revisions of this branch replaced the linear implementation in place; that made a brand-new system look like an evolution of the old one and put the default production path at risk. This revision restructures the work so that:

• v1 (linear trajectory) is preserved byte-exactly and stays the default. git diff against the pre-branch base is empty for every v1 module and test file.
• The tree implementation lands as an opt-in session server v2 (--use-session-server v2), fully contained in miles/rollout/session/v2/.

Commit series (atop the stacked picks)

1. test(session): nine byte-exact HTTP pins for the v1 retry/rollback surface, verified against the untouched tree — the default path can never move silently. Includes a characterization pin for v1's few-shot prompt-assistant anchoring quirk (out of v1's append-only + one-step-retry envelope; such flows belong to v2).
2. fix(tito): preserve-think kwargs become family constants on every template surface ("never cut think"), plus supports_midpath_assistant_rerender marking DeepSeek-V3.2's structural exception. Spike-verified byte-identical on pure tool loops.
3. feat(session): v2 trajectory tree — node = end of one model generation, full root→node token snapshots, pure deepest-prefix attach search, append-only commits under the session lock (MAX_NODES=1024 backstop).
4. feat(session): session server v2 — always-branch serving (mismatches grow siblings/new roots; retry semantics moves to the samples-op picker), --session-strict-append-only single-chain guard, compaction splice with zero-inheritance-root fallback, 409 on extending truncated generations, per-leaf assembly with default pick (temporal-supersession trim, roots never trimmed, non-retry-shaped trees 422) and default merge (exactly-once completion masking, rewards keyed by response id), both replaceable via --session-sample-picker-path / --session-merge-function-path. Version dispatch is a ~10-line seam in sessions.py; the codec is parameterized so the v1 wire stays byte-identical while v2 adds reward/metadata.
5. feat(rollout): driver-side seam — the agent's semantic layer (per-branch rewards) travels through collect_samples under v2 and comes back applied by the server-side merge; the v1 driver path (metadata overlays, {"max_seq_len": ...} collect body) is byte-equal to before, pinned.
6. test(session): tree-merge legality invariants — an independent oracle over randomized forests (seeded fuzz banks for retry shapes, truncation, guardrail-422 shapes) plus targeted edges (chained retries, equal-length twins, fork-below-fork ownership, mid-path length turns).

Testing

• v1: restored suites pass unmodified; A-list byte-exactness proven by empty git diff against base; new v1 pins green against the pristine tree.
• v2: HTTP matrix (branching, strict, truncation/compaction), session-state unit matrix (49), samples-op suite (16, incl. hook lanes), tree model (18), invariants fuzz (59). Router + session + generate_utils: 326 passed.
• chat-template suite: 1463 passed (1 pre-existing failure on upstream picks, unrelated).

Follow-ups

• GPU e2e for v2 branching flows (session_verify_runner branching-harness variant) on 4×H200 CI.
• v2 twins of the race-condition / pretokenized e2e suites (v1-only today).

🤖 Generated with Claude Code

https://claude.ai/code/session_01NmYsBAGK6NDt1LzjA7VSqH

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.