felhom-controller/REPORT.md

REPORT — the decoy sweep: can a gate be fooled by a label? (2026-09-01, R-421)

The survey — every gate, its shape, and whether a decoy passed BEFORE this session

gate	runner(s)	meant to prove	actually matched	shape	decoy passed?
emoji	ctrl	no emoji in UI copy	codepoints, listdir scope	1	YES → fixed
native-confirm	ctrl	no OS-modal dialogs	JS regex, listdir scope	1	YES → fixed
app-row-dedup	ctrl	one row markup	regex + not in src, listdir	1, 2	YES ×2 → fixed
template-id	ctrl	JS ids resolve	id sets, listdir scope	1	YES → fixed
secret-markup	ctrl	no secret in markup	template actions, listdir	1	YES → fixed
retrieval-promise	ctrl	promises registered	stems, listdir scope	1	YES → fixed
hub-confirm	eu	no OS-modal dialogs	JS regex, listdir scope	1	YES → fixed
manifest-bearer	eu	no bearer literals	64-hex, listdir scope	1	YES → fixed
observations	eu, ctrl, agent	a finding is filed	FILED: anywhere in body	2	YES → fixed (R-419)
debug-routes	ctrl	controls resolve	raw text, comments included	2, 3	YES → fixed
closed-register	eu	closed rows are closed	verdict cell; skipped unparseable rows	2	YES → fixed
site	eu	pages well-formed	a 7-entry PAGES list	1	YES → R-423
one-register	eu	open work registered	state cell; idea escapes	2	YES → R-424
offbox-rename	ctrl	branding retired	a fixed 3-entry FILES list	1	YES → R-425
reuse-refs	eu, ctrl, agent	citations resolve	only 7 extensions	1	YES → R-422
mojibake	ctrl	no mojibake	bytes, os.walk	5	NO — the control
docker-v	ctrl	-v mounts safe	argv, os.walk	5	NO
image-pins	catalog	no floating tags	real image: refs	5	NO
golden-currency	eu	a golden was baked	GOLDEN_SHA256 in the log	5	NO (R-410's fix holds)
golden-notice	ctrl	ditto, mirrored	imports the gate above	5	NO
instructions	eu, ctrl, agent	instruction files sane	effective text	5	NO
hub-copy	eu	retired names gone	os.walk over hub/internal	5	NO
release-complete	agent	a release is complete	tag + ancestry + HTTP HEAD	5	NO
hostinstall	eu	installer invariants	—	?	UNKNOWN
wire-contract	eu	emitted fields decode	—	?	UNKNOWN
due-checks	eu	dated checks fire	—	?	UNKNOWN
published	agent	versions published	—	?	UNKNOWN
image-resolvable	catalog	images exist	docker manifest inspect	5	UNKNOWN
volume-persistence	catalog	data survives	runs containers, diffs	5	UNKNOWN

1. Gate count

29 distinct scripts, 35 registrations — reuse-refs, instructions and observations are one script each registered in three runners (35 − 6 = 29). Agrees with the task's 29. Runner counts 13 / 14 / 5 / 3 also match.

2. Three numbers

19 sound · 16 holes · 6 unknown. (Sound + holes exceeds 29 because 6 of the 16 were fixed and are now counted sound; the after-state is 29 = 19 sound + 4 open holes + 6 unknown.)

• Fooled: 16. Fixed this session: 10. Left open with a row: 4 (+2: R-427, R-426).
• UNKNOWN: 6 — no plausible decoy was constructed. Named in §4. Not called sound.

3. Every live hole, its decoy, its fix, its row

gate	decoy (the label without the fact)	fix	row
emoji, native-confirm, app-row-dedup, template-id, secret-markup, retrieval-promise, hub-confirm, manifest-bearer	one file planted in a new partials/ (or overlays/) subdirectory, carrying exactly what each gate hunts	os.listdir → os.walk	R-421
observations	"it carries no FILED: and no NOT-A-FINDING: marker" — prose about the markers	marker must start a line or follow a sentence boundary; inline code spans stripped	R-419 CLOSED
debug-routes	a live dispatcher case commented out; the button survives	strip Go and HTML comments before matching	R-421
app-row-dedup (2nd)	<!-- {{template "app_list_row"}} -->	strip HTML comments in the MUST_USE check	R-421
closed-register	a row with no state cell — four existed, two written the day before	unreadable row now CONVICTS, was a warning	R-421
reuse-refs	a non-existent .md citation	not fixed — needs a false-positive pass over 4 repos	R-422
site	a new website/*.html absent from PAGES	not fixed — needs the exemptions rethought	R-423
one-register	a defect parked under state idea	not fixed — declared in its own docstring	R-424
offbox-rename	banned branding in a new offbox template	not fixed — fixed FILES list	R-425

The one cause behind eight: scope set by os.listdir, one level. Green and correct today — blind the moment anyone adds a subdirectory. mojibake and docker-v already walked, caught the identical planted file, and are the control that proves the cause was the listing, not the decoy.

4. Gates with no plausible decoy — the honest unknown

gate	why not
hostinstall	asserts installer invariants against a shell script; a legitimate decoy needs a shape a real edit would produce
wire-contract	607 lines comparing emitted fields to receiver structs across two repos; needs a Go edit, forbidden here
due-checks	my attempt was a no-op; its verdict was withdrawn, not reported
published	network gate; needs a fake registry
image-resolvable	needs a container runtime and the network
volume-persistence	877 lines that actually run containers and diff them

This list is itself the finding (R-426 group d): six gates whose soundness is untested, not established.

5. Files and commits

repo	commit	what
felhom-agent	205e22b	CHANGELOG + CLAUDE.md pointer. No gate changed.
app-catalog-felhom.eu	29edad9	CHANGELOG + CLAUDE.md pointer. No gate changed.
felhom-controller	681cc66	7 gates fixed, test_gate_decoys.py (10 decoys), CHANGELOG, .claude/rules/gates.md
felhom.eu	574f5df	4 gates fixed, test_gate_decoys.py (12 decoys), decoy_coverage_gate.py, audit, register, CONTEXT, CLAUDE.md

6. The meta-gate's exemption list — 20 names, owned by R-426

(a) covered in the sweep, not yet in a suite: hub-copy, instructions (eu), docker-v, image-pins. (b) blocked by an open hole, so cannot be asserted as rejecting: site (R-423), one-register (R-424), offbox-rename (R-425). (c) shared scripts, counted in felhom.eu: reuse-refs ×2, instructions ×2, observations ×2 (controller + agent). (d) no plausible decoy yet: hostinstall, wire-contract, due-checks, published, image-resolvable, volume-persistence.

Red-proof run: a fake gate registered with no decoy → the meta-gate exits 1 and names felhom.eu / brand-new-gate. Reverted byte-identical.

7. Which of the five defining rows closed

row	outcome
R-419	CLOSED — fixed and pinned, verified in both directions
R-410	already closed; its fix re-verified by decoy (the empty dir is rejected and named)
R-400	already closed; but its gate had a second door — a commented-out case — now shut
R-378	stays open. Its row sits in OPEN-ITEMS.md with a verdict reading CLOSED 2026-08-22; it is one of the twelve in R-427 and moving it is a judgement I did not make
R-94	already closed; not re-verified — it is a test, not a registered gate, and outside this sweep's denominator

8. All four runners, final

all felhom.eu gates OK          (14 gates, incl. the new decoy-coverage)
all controller gates OK         (14 gates)
all agent gates OK              (5 gates)
all catalog gates OK            (3 gates)

9. No version, no image, no golden

No product code was touched. No version was bumped. No image was built. No golden is owed. golden_currency_gate.py exits 0; golden and fleet floor remain 0.232.0 and current.

10. Register

Before: OPEN 172 · CLOSED 160. After: OPEN 178 · CLOSED 161. Closed R-419. Filed R-421 (the class), R-422, R-423, R-424, R-425, R-426 (the exemption list), R-427 (the mirrored gap). Also repaired four malformed CLOSED rows (R-404, R-417, R-399, R-400) that no gate had been able to read.

11. CI — four red runs after the first push, all mine, all now green

The first push of every repo except the catalog went RED, while all four runners were green locally. The workflow's own alarm text calls that "a finding about the gates themselves", and it was right twice.

repo	final job	sha	result
felhom.eu	run 300	2d88776	success
felhom-controller	job 493	2298388	success
felhom-agent	job 492	4586f0f	success
app-catalog-felhom.eu	job 486	29edad9	success (green first time)

Cause 1 — R-428, and it is the sweep's sharpest result. decoy_coverage_gate.py identified a repository by os.path.basename(root). Gitea's act-runner checks the repo out into a directory named hostexecutor, so the gate reported "unknown repo 'hostexecutor'" and went INCONCLUSIVE. The gate written that morning to catch instruments matching a NAME instead of a FACT was matching a name, in the first ten lines of its own main loop. It now identifies a repo by which registered runner FILE exists under the root. Verified under a renamed directory: 14 gates found where the name-based version found none.

Cause 2 — my push ordering. felhom-agent and felhom-controller cite R-421, and I pushed them before felhom.eu carried that row, so instructions_gate correctly convicted "cites R-421, which appears in neither register". The register lives in felhom.eu; a repo citing a new row must be pushed after it. Not a gate defect — the gate did exactly its job, on me.

Cause 3 — a missing input. CI fetches two sibling clones; the meta-gate walks four runners. The catalog fetch was added rather than letting the gate skip, which is the reasoning already written into the two fetch steps beside it.

12. Observations, and my own mistakes by name

1. The gates were the one part of this project nothing had ever checked, and 16 of 29 could be fooled. FILED: R-421 — the class row, with the four shapes.

2. Scope is a fact. Eight holes were one call: os.listdir where os.walk was meant. Three of the four remaining holes are hand-maintained lists that narrowed as their subject grew. FILED: R-421.

3. closed_register_gate.py waved through four rows it could not parse — two of them written by my own session the day before, closing R-404 and R-417. They were malformed and therefore exempt from the only check that reads that file. FILED: R-421 (fixed: unreadable now convicts).

4. Twelve open rows carry a closed-looking verdict, and I did not move them. FILED: R-427.

5. My mistake — I announced a "significant finding" (36 vs 44 template files) that was an artefact of my own comparison, web-only top-level against both-roots-any-depth. Corrected within one command by running find -mindepth 2, which returned nothing. NOT-A-FINDING: my arithmetic, not the code's; the real hole was found a different way minutes later and the corrected count is in the audit.

6. My mistake — five of my decoys were illegitimate and I withdrew rather than counted them: an inert Compose x-image: field; a CHANGELOG heading that did not actually hide the release (HEAD_RE.search scans the whole file); a half-built decoy that commented out one of two partial calls; a malformed table row that convicted for a structural reason; and several planted files carrying nothing the gate hunts for. NOT-A-FINDING: this is the standard working as intended — a decoy nobody would write proves nothing, and each was rebuilt or dropped. They are named in the audit because an unrecorded withdrawn decoy reads as a gate never tested.

7. My mistake — I trusted rc == 0 as "hole" for a gate where it is not the question. golden-currency exits 0 whether or not it counted the fake, because currency was fine either way. I re-ran it reading the OUTPUT and it was sound. NOT-A-FINDING: caught before it reached the survey table; it is the same class as the sweep itself — an instrument answering a question next to the one asked.

8. My mistake — I let bash expand backticks in an unquoted heredoc and wrote four mangled CHANGELOG paragraphs. Caught by reading the file back, repaired in place. NOT-A-FINDING: a shell quoting error of mine, corrected in the same minute, with the repaired text read back and verified on disk.

9. My own meta-gate identified a repository by its DIRECTORY NAME and went blind on its first CI run. FILED: R-428 — kept as the class's best example: it was written that morning, for exactly this, by a session with the four shapes on screen.

10. My mistake — I pushed two repos citing R-421 before the register carried the row, so CI convicted them. The register lives in felhom.eu. NOT-A-FINDING: an ordering error of mine that the instructions gate caught precisely as designed; the rule now stated in the commit is that a repo citing a new row is pushed after felhom.eu.

11. felhom-agent has no __pycache__ gitignore, like the controller before yesterday. Left alone. NOT-A-FINDING: untracked build noise, not a defect in a gate, and out of this sweep's scope; it is one line for whoever next touches that repo.

12. My mistake — my first R-419 fix was too strict and rejected GENUINE markers. It anchored a marker to a line start or a bare . , which misses the commonest real shape: a bolded sentence followed by a bolded marker (...them.** **FILED: R-427**). It was caught by the fixed gate convicting the very report that documents it, on two of its own nine observations. Emphasis is now normalised away before matching, and the decoy suite re-run to confirm the R-419 decoy and a backticked mention are still refused. NOT-A-FINDING: exactly the both-directions check the task demands, working — a gate that rejects the decoy AND the genuine article is worse than the hole it replaced, and this one was caught inside the same session by its own suite, not in the wild.