@@ -12166,3 +12166,6 @@ never opted in) biting six days after it was named.
 - **A release that is public while its legs are still running is a partial release the moment one leg fails** (subc-core-v0.17.4): the lane created the release non-draft, the index ingested it as each leg uploaded, the windows-x64 leg failed on a packaging path (`--target` puts the host build under `target/<triple>/` too), and the index then served "not yet published" for windows-x64 — honestly, and wrongly for the two main Windows users. CKCRED's draft gate held four times the same night for the same reason; it is now the lane's shape: draft at creation, a final job that needs every leg verifies the full asset inventory by name (5 tuples × 3 binaries × archive+sidecar) and undrafts. The index ignores drafts, so a failed leg leaves nothing an installer can see. Two-green-one-red pointed straight at the platform-specific path.
 - **Flake ledger — `provenance::tests::live_spawned_process_with_matching_start_time_still_matches` (linux-only)**: failed once on `ubuntu-latest` at 46dc1d00 (a docs-only commit; the same code passed at c31adaea and f03a5937). Reproduction attempt on the Ubuntu 24.04 arm64 VM: 0/40 idle, 0/30 under 4 busy loops on 2 cores. The assertion printed nothing but "not Match"; it now prints the variant so the next occurrence carries the discriminating fact. Not masked, not retried in code; the tip re-run is the verdict for the tag, and the ledger is the record.
 - **A convention held in one place and not another is indistinguishable from no convention to whoever reads only the second place** (CKCRED, fdd3f21): datum on stdout, note on stderr — followed for `mint-handle` two hundred lines above, not for `put`, whose last stdout line was advice. Only a caller capturing stdout notices, which is exactly who hit it. Same family as two sidecar producers with one fix. Also from the same fix: a patch anchor that matches zero leaves the build printing nothing, which reads as "compiles" — it was compiling the unmodified file.
+- **The tuple set lives in four places and the drives found each one missing in turn** (linux-arm64: index ✓, lane ✓, installer ✓, `ck` ✗; windows-arm64: lane ✓, installer ✓, index ✓, `ck` ✗): a release lane, the index worker, the bootstrap installers, and `AlphaTarget` each carry a copy, and only `ck`'s copy refuses at the point a user sees. The order the copies were found missing is the order they were added in, which is the tell — nothing binds the four. Owed: one source (the index's tuple set, or a fixture all four read) with a check that fails when any copy lags.
+- **A test that parses JSON tolerantly cannot see an encoding defect** (install.ps1 BOM): Windows PowerShell 5.1's `-Encoding UTF8` writes a byte-order mark; `ck`'s reader refuses it; Pester passed because `ConvertFrom-Json` tolerates it. Only a byte read (`first byte == 0x7B`) discriminates, proven by mutation on 5.1 (`Expected 0x7B, but got 239`).
+- **Served-script caches have two clocks** (raw.githubusercontent ~5 min, worker `max-age=300`): a pushed installer fix is live on `cortexkit.io/install` up to ~10 minutes later; a re-drive inside that window measures the old script. Check the served bytes for the discriminator before driving.