nx_research_journal_gate.nx

nx_research_journal_gate.nx -- the REFEREE for RES-R6 (crash-resumable journal). Proves nx_research_journal makes a research run RESUME from a crash point by REPLAYING the journal and SKIPPING already-completed phases -- AND proves the test can actually DETECT failure (the mandatory NEGATIVE CONTROL: the old, journal-less path re-runs everything, skipping nothing). GOOD lane (journaled): run 1 ("crash"): execute FETCH (3 sources) + phase_done, EXTRACT (3 sources) + phase_done, then VERIFY begin + 1 source_done, then STOP -- a crash MID-VERIFY (VERIFY never gets its phase_done). work_run1 counts every phase actually executed. run 2 ("restart"): rj_replay(GOOD) -> resume high-water = EXTRACT (VERIFY is NOT complete, so it is correctly NOT skipped). For FETCH..SYNTH: rj_should_skip -> FETCH+EXTRACT SKIPPED (counted), VERIFY+SYNTH executed. PROOF of resume = good_skipped >= 2 AND completed phases re-executed = 0. BAD lane (NEGATIVE CONTROL -- old path, NO journal): same crash, but restart has no journal -> rj_replay on the absent path returns 0 -> resume high-water 0 -> EVERY phase re-executes -> bad_skipped == 0 and bad_refetched == FULL_RUN. If bad_skipped != 0 the detector is worthless -> RED (no false green). TAMPER x2: (a) oversized record (rec_len+1 > cap=8) must be REJECTED with -2. (b) a TORN "RJ ... VERIFY ... DONE" fragment hand-written as a SEQUENCE of separate sys_write()s (no single-write framing) into a victim journal that ALSO holds a real FETCH+EXTRACT phase_done -> rj_replay must NOT let the torn VERIFY-DONE advance the resume point past EXTRACT. Proves a crash that tore the last line can never cause skipped-but-incomplete work. Output: rows -> stdout + evidence log knowledge/status/research_journal_gate.log, then a final verdict line. Exit 0 GREEN / 1 RED. Sovereign: only nx_syscalls + nx_framed_append + nx_research_journal. license_tier: ORIGINAL