finding(sideshow): findings 006 and 007 from the xorml harvest

_kos/findings/finding-006-doctor-census-trusted-a-manifest-that-travels-with-content.md

@@ -0,0 +1,30 @@
+# finding-006: doctor's content census trusted a manifest that travels with the content it checks
+
+**Date:** 2026-10-02
+**Subject:** how `sideshow doctor` decides an installed store version is intact
+**Occasion:** aae-orc-xorml half 2 (sideshow#138, #139) and aae-orc-mobz8
+**Evidence:** `internal/doctor/layer1.go` `checkContentCensus` at `c33fdd2`; an end-to-end run in a fresh
+store with a bmad 6.12.0 pack built after sideshow-packs#40; the operator ruling relayed on 2026-09-30.
+
+## Why this is worth writing down
+
+An integrity check is only as good as the reference it compares against. If the reference arrives with the
+content, replacing the content replaces the reference too, and the check passes.
+
+## What was observed
+
+- `store-content-census` reads bmad's own `_config/files-manifest.csv`. That file is part of the pack
+  content, so an `install --from` over an installed version (mobz8's case) brings or keeps a census that
+  agrees with whatever was copied. mobz8 recorded `store-content-census [ok]` after such an overwrite.
+- #139 added `store-file-manifest`, which recomputes every store file against the pipeline's
+  `file-manifest.csv` (inside the tarball since sideshow-packs#40). In a fresh store: a clean install reports
+  `2023 files match`; an overwrite from a copy with one edited `SKILL.md` and no manifest reports
+  `1 differ (first: .claude/skills/bmad-agent-architect/SKILL.md)` while `store-freeze` still reports ok;
+  the published r2 reports unavailable.
+
+## The remaining gap, and the ruling
+
+The pipeline manifest is still only as trustworthy as the install source: an overwrite from a tree that
+ships its own consistent `file-manifest.csv` is not detected. sideshow does not see the release signature
+or `install.meta` until install fetches and verifies signed tarballs (aae-orc-wk92). The operator ruled
+"yes default" on 2026-09-30: ship #139 as is, and anchoring to the signed copy waits on wk92.

_kos/findings/finding-007-install-never-deletes.md

@@ -0,0 +1,27 @@
+# finding-007: install copies over an existing store version and never deletes
+
+**Date:** 2026-10-02
+**Subject:** `InstallFromLocal` over an already-installed version
+**Occasion:** the same end-to-end run as finding-006
+**Evidence:** `internal/pack/pack.go` `InstallFromLocal` at `1b5d716` (a WalkDir copy with no removal step);
+the overwrite run in finding-006, where a manifest from the first install survived an overwrite from a tree
+that had none.
+
+## Why this is worth writing down
+
+A reinstall reads as "the store now holds this source". It does not: the store holds the union of every
+source installed at that version. That changes what a doctor finding means and what a reinstall can fix.
+
+## What was observed
+
+- The copy walks the source and writes each file; nothing removes store files the source does not have.
+- In finding-006's overwrite, the source had no `file-manifest.csv`, and doctor still read one: the file
+  from the earlier install had survived. That is why `store-file-manifest` caught the edit.
+- The same property means a file dropped between two builds of one version stays in the store after a
+  reinstall, and binding sync may keep serving it.
+
+## Open
+
+Whether a reinstall should replace the version directory (stage, then swap) rather than copy over it. That
+is install-path design, close to mobz8's refuse-or-force question, and is not decided here. doctor's
+`unlisted` count in `store-file-manifest` is the current way to see leftovers.

_kos/nodes/frontier/question-five-layer-doctor.yaml

@@ -31,6 +31,13 @@ content: |
   `aae-orc-e1jj` (the known-defects feed format) + `aae-orc-ztg5`
   (the registry itself).
 
+  Layer 1 progress (2026-10-02): finding-006 found the content census
+  trusted bmad's own manifest, which an overwrite replaces; #139 added
+  store-file-manifest against the pipeline manifest, and anchoring it to
+  the signed release waits on aae-orc-wk92 (operator ruling). finding-007:
+  install never deletes, so a store version holds the union of every
+  source installed at that version.
+
   Open sub-questions:
   - Layer 2 spec separately or fold into the weaving engine?
   - When should layer 3 escalate from warn to error?