Deliverable recoveries/helm-root-redirect.md (12,811 B / sha256
e68be4cdd5a535e22eae87f5ffbf9cb6841ea3d8a044ae34c13e74f86460568b). Scratch tree
/root/crucible/q44/; harness run_negatives.py -> out/Q44-NEGATIVES.json
(66,881 B / 611ab1e06e172be535ae5f19e21b4e1a6667a0f1b7542d130fc9d72298c5b895,
generated_at 2026-10-12T20:49:04Z), 17 invocations. Nothing deployed, no cp into
/root/.helm/bin/, decisions/ and inbox/ untouched, no another-party file edited,
no secret.
The pair: subject collector 4e45021d80395d3f... (29,954 B / 720 lines) byte-identical
to Q43's -- every change is in the gate. Gate 18580c102ddb1c3f...
(94,716 B / 1,731 lines) against Q43's c22601ecea4192d9... (88,468 / 1,612);
diff -u = 195 lines = 3 header + 133 added + 14 removed + 45 context, and
1,612 + 133 - 14 = 1,731 exactly.
What changed: the copy builder rewrites the HELM_HOME = assignment ITSELF to the
per-invocation sandbox, so every constant composed from it follows by construction;
the composed set is found by ast on the subject's own source (module-level Assign
nodes whose value references HELM_HOME) and is [SNAPSHOT_PATH, HELD_DIR]. The
SNAPSHOT_PATH rewrite is kept as defence in depth. Rule (3)'s --pin refusal is kept
but reclassified as defence in depth. Q39's admission, its behavioural probe and its
per-leg watch are untouched; no check's meaning changed. Three new checks, so the
green count is Q43's 166 + 3 = 169 passed / 0 failed: (1) the root constant is
redirected; (2) every constant composed from the root resolves inside the sandbox,
measured by a subprocess that imports the copy and prints the live values (not a
source marker); (3) the live held directory is unchanged by every leg, censused on
bytes, sha256, mtime_ns and ctime_ns before any leg and again at the foot.
Green pair, 5 consecutive invocations: rc 0 each, 12,966 B stdout, "helm-selftest:
169 passed, 0 failed" / VERDICT: ALL GATES PASS, 12.00-12.14 s; log digests
5c519157eaec4a40... / 97916ab27b8521c0... / 6f5553364b7a0d09... / ad57bad15689dfc5...
/ 00a1ae6dc2519efc... . Q43's counters unchanged (4 legs redirected, 3 --dry-run legs
original, 0 refusals, sandbox holds 3 paths with no data/snapshot.json).
The five Q39 negatives, still refused -- each run under the Q44 gate and, as the
pre-Q44 comparator, under the Q43 gate: all ten rc 2, stdout 0 B
(e3b0c44298fc1c149...), no report block, live delta 0, live held delta 0, 0.14-0.21 s.
The discriminator, on a scratch protected root (never live), rule (3) disarmed and a
--pin --deploy leg appended at the foot so it runs through the same leg wrapper.
pinleg (redirect armed): rc 0 / 169 passed, the pin lands in the sandbox
("pinned e40308e573d046d7d503acecbfbde82b8ccd10758a345b0440d55b796e2d821e ->
<sandbox>/data/held/stackchain-home.html.sha256"), protected held file delta 0
(114 B / b58d2d0eb9f45862b... / mtime_ns 1791478687000000000 / ctime_ns
1791838054542473495 before and after). pinleg-noredirect (redirect disarmed, live root
stood in for the scratch root): rc 1 / 166 passed / 3 failed, the same shape lands in
the protected root's held file (120 B / f98215b460bf89764924096c6fc657e62173e4c26f99bc4a2e5c489f3ed0f416 /
mtime_ns 1791838143983952745 / ctime_ns 1791838143991953414 -- NONZERO). Prevention
separated from detection by measurement. The live-path write was NOT re-measured: it
would clobber /root/.helm/data/held, Q43 already measured it (pinned 5e5cd6d26a70...)
and this run's pinleg reproduces the same body digest e40308e5... from the same code
path, so the class is the same one.
Live state, delta 0 across all 17 invocations: snapshot 14,039 B /
aa842b0fd88037f83115b4a54d617f27666f4de06ef84f9a64c8fd0a5f04656d / mtime_ns
1791479579475780303 / ctime_ns 1791789490146606356; held pin 120 B /
5b6d70716153c3f4bd82045d5d71edf88f793a240ce1275c3dfca5759d0ee297 / mtime_ns
1791478687099553071 / ctime_ns 1791478687102553111. Deployed bin untouched
(ee0b36a82f23fcf3... / 7ed1bca8f9da624f... / 5c67473d89283c3d...).
Limits and this run's own defects, logged first in the deliverable: the redirect keys
on the HELM_HOME = assignment, so a differently-named root or a call-time path is not
covered by the substitution (Q39's literal scan and the per-leg watch remain the
backstop); the shipped gate never exercises the redirect on a --pin leg, so that
shape's evidence is the variant; my first green run read 168 passed / 1 failed because
my own comparator excluded the sandbox root itself (startswith(dir + "/") is false for
HELM_HOME, whose value IS the root) -- fixed in patch_gate2.py and the only reason the
count is 169; the Q43 resolution-leg check false-fails under the variants and was fixed
in the variant only (patch_variants.py), the shipped gate untouched by that edit; and
one unexplained observation -- in pinleg-noredirect the sandbox census holds 5 paths
including data/snapshot.json, i.e. a leg wrote a snapshot inside the sandbox, which is
the sandbox doing its job and rides on nothing, recorded rather than explained.
Q7, measured: after this item the Helm lane's executable set is empty again (Q9 blocked
on the audio handover, Q13/Q19 gated on Alexander, Q42 open and zephyrus's). No new item
opened. The section 4 adoption is now a three-file ask: collector 4e45021d... + gate
18580c1..., with Q40's finding unchanged -- the deployed gate 7ed1bca8... has no
subject-selection path at all.
Merlin
534be9e32c