Machine-check the post-emit read-set of the 15 audit-unread hosted-template fields (no new kappa)

notes/agent-45-task-card.md

@@ -0,0 +1,70 @@
+# Local task card — agent-45 (written before substantial computation)
+
+Task ID / proposed scope key: `uniform-transfer-audit` (chooser scope key `603f341b99f590f17a76`),
+refined to the hosted-template post-emit field-contract read-set audit.
+Track: adversarial-audit. Chooser nonce: `2bb90ec128a184bdec47e263785b9c5d` (OS-random draw,
+priority revision 4, catalogue SHA-256 `d46eb6d5b3e279428753d770653663fbdf1063e55c255ea073d768f344927ef7`).
+
+Base commit; exact predecessor and dependency heads:
+- Base main `0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2`; selected-record SHA-256 `10b9e340fd8e354a37a0c4813c90e5de1dd42044b5f6aca4a3fc5a42b685fa32`.
+- Predecessor: Josie Elliston's PR1959 `research/t16-audit-mutant-coverage` (on main at base),
+  whose README records the exact claim audited here and labels it "a text search, not a proof".
+- Adjacent concurrent work: PR2158 `research/invoice-mutant-coverage` (invoice `run()` mutation
+  testing; different object and method — dynamic mutation of invoice predicates vs static
+  read-set/whole-object channel audit across all pinned post-emit consumers).
+
+Concrete object and question:
+The frozen PR1087 `code/audit_allcold.py` acceptance audit does not read 15 hosted-template
+fields (`cover`, `m`, `version`, `port_restriction`, `provider_dispatch`, `replica_lift`,
+`route_semantics`, `supplier_substitution`, `expanded_scalar_interface`,
+`hosting.{rule, scope, idle_remainder_semantics, owner_map_sha256}`,
+`async_schedule.{scope, source_plan_sha256}`). PR1959 asserts by text search that no stage
+after `emit_allcold.py` reads any of them. Question: is that claim machine-checkable over the
+pinned consumer universe (frozen PR1087 archive `code/` + maintainer frontier1094 scripts),
+and through which whole-object channels (hash binding, serialization, iteration, equality)
+could a change to those 15 fields still reach a post-emit acceptance decision?
+
+Scope status: refined; reason: the card asks for one exact hosting/transfer implication with
+explicit quantified hypotheses; the refined object is one named claim ("no post-emit reader of
+the 15 audit-unread fields") whose downstream consequence is the template acceptance boundary.
+
+Existing restriction being changed (or exact claim being audited):
+Exact claim audited: research/t16-audit-mutant-coverage/README.md, "Accepted: 19 path mutations"
+bullet ("A text search of the PR1087 archive and the maintainer frontier1094 scripts finds no
+reader of the 15 unread fields in the hosted template after `emit`. This is a text search, not
+a proof."). No restriction is changed; the audit upgrades the evidence class.
+
+Method; explicit parameter ranges/candidate alphabet:
+stdlib-only AST read-set analysis (Python `ast`): every constant string equal to one of the 15
+names in every pinned file is classified by syntactic context (subscript key, attribute, call
+argument of `get`/`pop`/`setdefault`, plain string, other); every dynamic-access site
+(`x[name]`, `getattr(x, name)`, `**` unpacking, whole-object iteration/serialization/hash/
+comparison) is enumerated per file. Universe: all `.py` members of
+`research/complex-t16-rare-pattern-exchanges/upstream1087.tar.gz` (SHA-256 pinned) plus
+`scripts/verify_frontier1094*.py` and `research/community-batch1099-audit/check_frontier1094.py`
+and the selected stage wrapper `research/complex-t16-rare-pattern-exchanges/verify.py`.
+Falsification first: any name-directed post-emit read or unresolvable dynamic access falsifies
+the claim at the smallest instance (the exact file/line).
+
+Nearest overlapping PRs and why this adds something:
+- PR1959 (merged): raised the gap and left it as a text search; open question 2 is a different
+  (driver-binding) item blocked on a maintainer decision. This adds the missing machine check.
+- PR2158 (open): dynamic mutation of `invoice_allcold.py` predicates; mine is a static,
+  universe-complete channel map over all post-emit consumers, and states exactly which channels
+  mutation testing can and cannot reach.
+
+Optimistic full cost budget; required observable improvement:
+No kappa/supplier claim. Deliverable is a checked statement + channel map with pinned hashes;
+cost is bounded by universe size (a few dozen source files); runtime target well under 60 s.
+
+Success evidence; falsification test; time/compute cap:
+Success: `verify.py` prints PASS with EXPECTED.json equality over the pinned universe; a
+corruption control (a name-directed read injected into a scratch copy) flips the result.
+Falsification: a single name-directed post-emit read of any of the 15 names, or a dynamic access
+that cannot be resolved from constants. Compute cap: seconds per run; one bounded pivot maximum.
+
+Predecessor authors; intended artifact and reproduction command:
+Josie Elliston (PR1959); Jeremy Kothe (PR1087/PR870 frozen code); Rohan Arun (PR1094 selected
+run); agent-43/Aaron Wang (PR2158, adjacent). Intended artifact: `research/hosted-template-postemit-fields/`
+(README, TASK.json, SOURCES.json, VERIFY.json, verify.py, EXPECTED.json). Reproduce:
+`python3 -B research/hosted-template-postemit-fields/verify.py`.