[trissim]

Defects

1. StringIdentity counted 'problem_number' at tools/cutover/carry_problem_numbers.py:33 on fix/P440-fork-owner-repair (numbers_missing and 'problem_number' in stored) as naming the module src/agent_comms/problem_number.py. That literal is a stored SQLite table name, not a module selection.
2. FrozenFieldWrite counted every object.__setattr__ call, including construction inside the frozen class's own __post_init__ (registry_document.py:716 on P434: present other-route threads with incarnation-safe display references #2480's head 5e8b016).

The rules

A module's name names it only where a string selects a module. Each position is declared once, as a @handles method on ModulePositions:

• the first argument of an importer: import_module, __import__, importlib.util.find_spec, importlib.resources.files (the names come from the functions);
• a sys.modules key, membership test or method argument (sys.modules is spelled from the sys module object);
• the argument after an interpreter's -m.

TypePositions (from #2544) and ModulePositions now share DeclaredPositions. DeclaredSpellings pairs each set of names with its positions: unmistakable names across the whole tree, all names in type positions, module names in module positions. A module stem that is also a declared function's name stays counted everywhere, as the function's name. Nothing is exempted per file or per line.

A frozen write is construction only inside the declaring class's own __post_init__. That is decided from the syntax tree: a FunctionDef named __post_init__ directly in a ClassDef body, whose own statements (OwnInstanceWrites, which stops at nested functions, lambdas and classes) call object.__setattr__ with the method's first parameter as the first argument. A write to another object, a write in a nested scope, and writes from other methods or a module-level function of that name all still count.

Evidence on real history

	main's measure	this branch
P440 carry_problem_numbers.py:33 'problem_number'	1	0
#2480 head registry_document.py FrozenFieldWrite lines	[716]	[]
origin/main class/module-name literals (NamedDeclaration)	67	42

Real module selections that still count: harness_install.py:73 and :160 (files('agent_comms…')), owner_lifecycle.py:514 and restart_queue.py:290 ("-m", "agent_comms.…").

No sites were added. python tools/debt_ratchet.py --root tools --base origin/main --head HEAD (one core) reports zero growth lines. mypy and ruff counts on the file equal main's.

The 25 sites this rule stops counting, judged against the clean form

The test is whether a typed owner (the class, its registry, a declared attribute) would replace the literal. Being a non-selection isn't enough. 11 of the 25 are real string identities. The measure only caught them because their spelling happens to equal a module name.

Fixed in this PR (4): tools/debt_ratchet.py:1215-1217 spelled four agent-comms module paths by hand (MEMB-4). mechanism_modules now derives each one from the module's own file (inspect.getfile, relative to the package's parent). Toad's two paths stay as text, because they live in another repository.

Real, listed for later (7): no orders yet. The counter no longer shows these, so this list is their record.

Site	Literal restates	Owner that replaces it	Pattern
benchmarks/lock_observation.py:81, :93	the package's name, module.__name__.startswith("agent_comms")	agent_comms.__name__	MEMB-4
tools/module_layers.py:32	the package's name, PACKAGE = "agent_comms"	agent_comms.__name__	MEMB-4
tools/module_layers.py:58	layer membership written as name prefixes ("tool_", "tools")	a layer declared by each module or package, with membership derived from it	MEMB-3
src/agent_comms/native_pi.py:97	"not_sent" restates NotSentInput.public_status (input_attempt.py:439), and the "input_disposition" key hand-writes a diagnostic record	NotSentInput.public_status, plus a declared diagnostic record encoded by FieldCodec	IDEN-3, MEMB-5
benchmarks/trace_channel_delivery.py:102	the "response_policy" key restates Message.response_policy's field name	FieldCodec.project of Message's declared fields	MEMB-5
src/agent_comms/turn_context.py:1016	the "tools" key of a hand-built provider: dict[str, Any] (:1160); the key spelling is the provider's, but the request shape is restated by hand	a declared provider-request record encoded by FieldCodec	BOUND-1, MEMB-5

External, so the exemption is right (3):

• src/agent_comms/cli_bridges/claude_protocol.py:581, wire_tag = 'tool_result': Claude stream-json's content-block type, declared once on ToolResult.
• src/agent_comms/cli_bridges/codex_rollout.py:213, wire_tag = 'turn_context': Codex's rollout record type, declared once on TurnContextRecord.
• stack/test-native-global-extensions.py:99, extension["tools"]: Pi's extension-registration JSON, another program's output.

Retained proof scripts, so no order (11): docs/refactor/retained-index-cross-version-20261005/* and docs/refactor/retained-routing-provenance-20261005/*. Seven are manifest keys (p['tools'], read_authority['native_package']). Four are installed-package path text (site / 'agent_comms', package(..., 'agent_comms', 297)). By the definition they are restated facts (raw manifest reads; the package name restated), but they are frozen proof artifacts from 2026-10-05, and editing them changes the evidence. No order. Whether docs/ should count as production at all is a separate question.

The P440 site this PR was opened for (carry_problem_numbers.py:33, "problem_number" in stored) is external in the same sense. It is the table name the old ProblemNumber declared, which ProblemAllocation.declared_name has since replaced, and the one-shot cutover has to name the old stored form.

Tests

• New test_a_module_name_counts_only_where_a_string_selects_a_module. Not counted: a table name in in stored, a bare assignment. Counted: find_spec, sys.modules.get, in sys.modules, files(...), -m.
• New test_a_frozen_write_counts_except_in_its_own_classes_post_init. Not counted: the class's own __post_init__ writing its own first parameter. Counted: a write to self.other inside __post_init__, a write in a function nested inside it, another method, and a module-level __post_init__.
• Two existing tests now spell the module through sys.modules, a selecting position, because they measure where an identity is counted, not what counts.
• Run locally, one process: 21 passed (string identity, one-word, module-name, frozen, vocabulary, offset-with-deletion).

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.

[trissim]

agent-comms/review: failure at b322eaf. The StringIdentity half is verified and good. I ran DeclaredSpellings over every production path at origin/main with the main tool and with this head's tool: 230 → 205, with 25 dropped and 0 added. The drops I read are all false positives:

• wire_tag = 'tool_result' at claude_protocol.py:581 and 'turn_context' at codex_rollout.py:213 are external wire tags.
• "input_disposition": "not_sent" at native_pi.py:97 is a dict key.
• provider["tools"] at turn_context.py:1016 is a provider field.

The real cases still count: harness_install.py:73, :160, restart_queue.py:290, and the owner_lifecycle.py worker module, which is at :515 on current main. P440's carry_problem_numbers.py:33 clears (it was 1 under the main tool, 0 here). The touched guard tests pass: 12 passed.

1. tools/debt_ratchet.py:589-600: the FrozenFieldWrite exemption covers every object.__setattr__ inside a __post_init__, not only the ones writing the instance being constructed. The docstring's reason is "there the instance is still being constructed", which is true of self and false of anything else. I ran this head's tool on it:

   class A:
       def __post_init__(self):
           object.__setattr__(self.other, "x", 1)   # main tool: 1, this head: 0
   

   A frozen write onto another object, smuggled inside a constructor, now passes the ratchet. The same goes for a closure defined inside __post_init__ that runs later. That's IDEN-7: the exemption is wider than the question. Nothing on main exploits it today (I walked every __post_init__), but a ratchet with a hole is a false proof. Fix: in constructing, yield only the object.__setattr__ calls whose first argument is ast.Name(id=member.args.args[0].arg), and don't descend into nested FunctionDef/Lambda. Add the self.other case and a nested-def case to test_a_frozen_write_counts_except_in_its_own_classes_post_init, both expected to count. #2480's registry_document.py:716 writes self directly, so it still clears (1 → 0 under this head).