session 2026-07-18 S4: prep→fleet + batch 2 graduated (table credit real); unrecognized census+positive-test; S4 session file + 3 KG lines + 1 harvest proposal

claude/memory/knowledge-graph.jsonl

@@ -390,3 +390,6 @@
 {"subject":"claude-code","predicate":"drift-pattern","object":"mispinned-sandbox-runs-against-reality — the PENDING-63 ratification test monkeypatched a FRESH module object (test harness _load() = spec_from_file_location, never sys.modules) while the code-under-test (source_gate) imports via sys.modules: the gate-driven leg silently ran against the REAL Chamber Sources archive from inside the fleet. Only the refusal path kept it read-only; a matching canonical_slug in the real pending-manifest would have let a TEST mutate the steward's archive. Antidote: pin sandbox patches on the SAME module object the driven code imports, AND census the real substrate before+after any test that can touch a shared store. The witnessed red DETECTED it (S2's watched-red discipline earning its instance). Harvest: sandbox-must-pin-the-shared-module (ladder, PROPOSED).","valid_from":"2026-07-18","valid_to":null,"confidence":1.0,"source_file":"session-2026-07-18-s3-first-loeb-graduations-pending63-closed.md","extracted_at":"2026-07-18"}
 {"subject":"claude-code","predicate":"drift-pattern","object":"truncating-a-narrating-stream-mutates-the-run — piped a prints-as-it-writes producer through `| head -80`; head's exit SIGPIPE-killed PRODUCTION at 77/144 works with the death invisible (exit swallowed by the pipe). Census-through-truncation upgraded: here the truncation was a MUTATING act, not just a partial view. Antidote: never truncate a stream that narrates a run — tee to a log and read it whole; check pipestatus. Self-caught same session by re-run + tail.","valid_from":"2026-07-18","valid_to":null,"confidence":1.0,"source_file":"session-2026-07-18-s3-first-loeb-graduations-pending63-closed.md","extracted_at":"2026-07-18"}
 {"subject":"chamber-library","predicate":"state-change","object":"first 6 Loeb graduated 2026-07-18 (REVIEWED-58 + observed REVIEWED-63 apply); compound-source archive form LIVE (loeb-dsl-537-volumes artifact owns the hash; works attach by id); pending→archive seam repaired; Block 1 closed; canon 1300; fleet 157/157","valid_from":"2026-07-18","valid_to":null,"confidence":1.0,"source_file":"session-2026-07-18-s3-first-loeb-graduations-pending63-closed.md","extracted_at":"2026-07-18"}
+{"subject":"claude-code","predicate":"drift-pattern","object":"silent-non-firing-read-as-reassurance — reported `unrecognized: 0 across 11 real works` as if informative; steward named it uninformative (the negative-grep class, generalized from bypass properties to safety-net silence) AND named the failure no gate can see: a flattened structure keeps its words (fab 0 · lost 0 · green) while the TYPING silently dies — preserve-never-flatten invisible to body-conservation. Antidote: CENSUS the substrate for what WOULD fire (silence becomes informative only against a censused absence of triggers), then positive-test at a censused trigger; keep the INVERTED watch-list (works where the net MUST fire — a Pausanias batch reporting 0 is a broken net, not a clean book). Closed same hour: DSL census (17 novel-structure works) + real firing (Menander/Pausanias/NE).","valid_from":"2026-07-18","valid_to":null,"confidence":1.0,"source_file":"session-2026-07-18-s4-prep-fleet-batch2-unrecognized-proven.md","extracted_at":"2026-07-18"}
+{"subject":"claude-code","predicate":"drift-pattern","object":"canary-clean-over-standing-failures — 'engine canary clean' phrasing over a gate carrying 5 known-FAILED rows (committed state since 07-10); the honest claim was 'coverage-ledger byte-identical: NO DELTA from this graduation'. Self-caught at phrase level (substrate-diffed before asserting); steward deepened the design consequence: a gate with standing known-failures STOPS GATING those rows — 'is there a NEW failure' becomes undetectable without a remembered baseline; known-failing and newly-failing must be structurally distinguishable (the ABSTAIN/UNVERIFIED distinction at the gate-reader layer). Antidote: state deltas against the committed baseline, never cleanliness; design item carried to the engine session.","valid_from":"2026-07-18","valid_to":null,"confidence":1.0,"source_file":"session-2026-07-18-s4-prep-fleet-batch2-unrecognized-proven.md","extracted_at":"2026-07-18"}
+{"subject":"claude-code","predicate":"drift-pattern","object":"calibration (antidotes self-firing): witnessed-red DISCHARGED on first recurrence (fleet observed failing 157/158 before the tool existed — the S2 fixtures-first honesty note closed); scratchpad-volatility antidote self-fired at the point of work (census JSON+script moved INTO the repo before commit — the S3 prep-script trap not repeated); relay-precision held in BOTH directions (steward's 'override branch never fired' was half-right: author_overrides WAS fixtured, lang was not — corrected precisely rather than absorbed).","valid_from":"2026-07-18","valid_to":null,"confidence":1.0,"source_file":"session-2026-07-18-s4-prep-fleet-batch2-unrecognized-proven.md","extracted_at":"2026-07-18"}