|
| 1 | +# Round 28c — C3 (rc.5 #1) and Y2 (alpha.3 ×2) |
| 2 | + |
| 3 | +No keys here (redacted CI artifacts only). Offline re-check: `ci-check.py check` on |
| 4 | +the downloaded `dsh-e2e-<run>` artifact, with the 300-char detail cut disabled. |
| 5 | + |
| 6 | +## Y2 on DSH 0.2.2-alpha.3 — 已复现 2/2 (41842392546, 41842397847); 0 on rc.5 |
| 7 | + |
| 8 | +Y2 asserts that in y-2 (new process after kill -9 during a historian hold) the |
| 9 | +**first step's** busy probe reads false. On alpha.3 the first step's status dump |
| 10 | +(41842392546) is: |
| 11 | + |
| 12 | + killedPid 4789 (y-1) -> y-2 first step: historianBusy=true, |
| 13 | + leases=[{holder dsh-historian:…:4847:a8cca8e6, owner_pid 4847, expires_at +300 s}] |
| 14 | + lease acquired at …153217, step status at …153221 (4 ms later) |
| 15 | + |
| 16 | +pid 4847 is the NEW y-2 process, not the dead holder: the dead lease was taken |
| 17 | +over immediately (no 5-min TTL wait), both y-2 historian passes ran `noop`, and |
| 18 | +after the turn there is no live lease and `compartment_in_progress = 0`. Every |
| 19 | +other Y2 condition holds; `still_dead` is empty. On rc.5 (41839150646) the same |
| 20 | +pre-step trigger fires, but by the step's status dump the noop pass has already |
| 21 | +released the lease (`leases: []`, busy false). The y-2 log order is identical on |
| 22 | +both versions (trigger → native check → step_start). |
| 23 | + |
| 24 | +Classification: **timing assumption of the Y2 check** (the probe samples while |
| 25 | +the new process's own pass still holds its fresh lease); Magic's behaviour |
| 26 | +(dead-holder takeover, no wait, nothing left behind) is correct on alpha.3. |
| 27 | +Why alpha.3 starts the step sooner is **static inference** (faster pre-step on |
| 28 | +alpha.3), not reproduced locally. Not changed: the check is not relaxed (user |
| 29 | +rule); whether Y2 should accept a busy probe whose only lease belongs to the |
| 30 | +live new process is listed for the user. |
| 31 | + |
| 32 | +## C3 on rc.5 #1 (41842386831) — first C3 FAIL in the 15 runs re-checked from round 27 on |
| 33 | + |
| 34 | +C3: the lowest prompt-token value **after the global peak** must be below it. |
| 35 | + |
| 36 | + cmp-02 [10313, 15093, 19008, 19084] cmp-03 [19301, 24067, 19235, 19385] |
| 37 | + cmp-04 [19578, 23206, 23364, 23729] cmp-05 [23923, 22564] |
| 38 | + cmp-06 [21115, 27264, 27777] cmp-q [26262] |
| 39 | + use-search [26271, 26504] use-memory [27460, 27536] use-expand [28351, 28638] |
| 40 | + peak 28638 (use-expand, the LAST step) -> nothing after it -> FAIL |
| 41 | + |
| 42 | +Context did drop (cmp-03 24067 → 19235, cmp-05 → 22564, cmp-06 21115), the |
| 43 | +historian fired (C1), wrote 7 contiguous compartments (C2/C6), the codeword was |
| 44 | +answered (C4/C8). Compartments first appeared after cmp-05 (the e2e |
| 45 | +corrupt-once injection + repair retry landed there; on 41839150646 after cmp-04), |
| 46 | +so the scenario used its 6th read pair; the two newest full-file reads stayed in |
| 47 | +the protected tail and the U turns ran at 66–72 %, each adding a little, so the |
| 48 | +peak is the final step. Classification: **待验证** — most likely scenario |
| 49 | +variance plus a check that assumes the peak precedes compression; no adapter |
| 50 | +fault found in the logs (I2 PASS, no re-reads; the model once misread a |
| 51 | +one-call fold note as "both files ran" and then read the second file, which had |
| 52 | +not been read yet — correct new work). Check unchanged. |
0 commit comments