[michaellchung]

Summary

Fixes found during the S23 third-stage GPU runs (full-parameter, SAO streaming full-parameter, and P0-S), on top of #257. The tree at the head of this branch is identical to the integration branch s23-final2.

What changed

Full-parameter islands:

• 34ea9907 Lay out a single-role full-parameter stream without a role lane (defect 3: dense role stream).
• a2f4ce76, e0fb3d4f, 259e4bce Bind reward-function trajectory evidence for non-grader full-parameter islands (defect 4: dense evidence kind was hard-coded to secrlenv), with the spec update in rl-drop-reward-tamper-hash.
• 9de5e7db Install the actor step recorders for a full-parameter trainer (defect 5: no per-step LR record).
• 4c636f8c Report initial-cut upload progress and mirror trajectory evidence to the tape volume.

SAO streaming full-parameter:

• 5ddc0bd9 Print the observed full-parameter layout as an attestation record on refusal.
• aacd367e Allow the critic-family mechanisms under sao-streaming-full (defect 6).
• 120e8548 Reload the SAO contract bundle on the head controller (defect 7).
• 08599f3c Wait for the initial cut off the event loop (defect 8: two-island deadlock).
• f3ca2460 Drop the removed SAO allowance fixture from the real-syncer test import (the suite failed at collection after 120e8548).

P0-S:

• 6e35b232 Tape each accepted group's generating policy under a policy-age limit (rl_group_policies event).
• 3f42ec47 Resume a round cut from the data cursor the cut recorded (fully_async has no ledger data cursor).

GPU evidence (summary; Modal, cost figures are estimates)

• Full-parameter run b: 3 rounds, rc=0. Versions 0-3 publish the same hash on both islands; schema-4 evidence 6/6 matches behavior version and hash; defect 5 fix is effective. Run c (sync init timeout, suspected l0-to-head transfer) root cause is not verified.
• SAO streaming full-parameter main-20261014d: rc=0, 1 round x 2 shards; all five criteria passed on manual review. Defect 9 (Miles critic port bind race, EADDRINUSE) did not reproduce and is not fixed.
• P0-S: run a (3 rounds + cut) passed; run b2 resumed from the cut for 3 rounds: 23 groups returned, all regenerated, no stale trajectory trained; stale-drop count never negative; carry-over execution mode partitioned-overlap; each group's policy hash equals the published hash of that version.

Tests

Full local suite (with cargo) on the same tree (s23-final2): 5649 passed, 0 failed, 44 skipped. The 7 multinode thread-limit tests in the baseline passed this time (about 2,250 user threads).

Known gaps

• rl_round_trained label/execution_mode still reads partitioned-serial for fully_async rounds.
• The fully_async ledger does not record a data cursor; restarts not from a cut are not verified.

Depends on

#257.

🤖 Generated with Claude Code