Close the receipt stack's outstanding review Lows

[MaxGhenis]

Follow-up to #571, #579 and #588, the receipt stack now on main. It collects the Lows their independent reviews left as non-blocking.

Change

False refusal, fixed (#571 round 16). A helper annotated List['Annotated[float, "weekly | monthly"]'] inside a measure function was refused, while the unquoted form bound. The typing-alias check reparsed the forward reference and read the Annotated metadata string as a type. holds_union now walks type expressions structurally:

• it reads only the first element of Annotated;
• it never parses Literal values;
• it handles quoted forward references the same way.

Documented limitation (#571 round 16, narrowed after this PR's round 1). Any use of id or another refused name (globals, eval, …) as a variable is refused: a read, an assignment target or a del, even of a parameter or local, such as reading a household identifier named id. Telling a local from the builtin needs lexical-scope analysis, which this rule deliberately doesn't attempt. Positions that are not variable uses are allowed: an unread parameter or function of that spelling, an attribute (simulation.id) and a keyword argument (id=3). The docs now say exactly that and suggest renaming; a new test pins each position.

Round 1 of this PR's review (GPT-6.1 Sol, REQUEST-CHANGES) found the structural walk's Annotated exemption applied when any binding of the name was Annotated. A helper that rebinds it (from typing import Tuple as Annotated) hid a real union: the reviewer's end-to-end case bound and produced 29 or 28 pennies depending on typing's cache. Fixed in a15a98f:

• the Literal and Annotated exemptions now need every possible binding to be that form, in both the marking pass and the walk; an ambiguous name is inspected as a type, and the reviewer's case is refused;
• the walk and name resolution are iterative, so quoted layers deeper than the recursion limit no longer raise RecursionError; the reviewer's 32-case deep-quote probe now matches main on every case.

Four mutants (each exemption back to any-binding, and the recursive walk) are each killed by the new tests.

Regressions for surviving mutants and untested paths:

From	Test	Kills
#571 r16	List[Optional[Decimal]] inside a function is refused	dropping Optional from the union forms
#571 r16	a certificate naming other bytes gives an uncertified receipt, not a validation error	dropping the certified-bytes comparison
#571 r16	quoted and unquoted Annotated metadata with | binds	reading all of Annotated's elements
#571 r19	fresh process: Final['_receipt_forward_probe'] poisoned in typing's cache before anything binds is refused (the test asserts the module really received the poisoned alias)	— (pins the interpreter-constant flag)
#579 r5	a stale composition member is caught first or last	dropping the member re-check
#579 r5	a current-law member that later gains changes is caught, alone, mixed and nested	dropping the current-law composition's handle
#588 r4	a str subclass with its own __str__ hashes as its value	str(value) instead of str.__str__(value)

Round 16 had also reported a surviving mutant that deletes module bindings in the alias resolver. The existing tests kill it at main's head (18 failures), so nothing was needed.

Tests

• Receipt suites, editable layout (--noconftest): 936 passed, 2 skipped (round 1 added 11 tests).
• Receipt suites plus the budget contract, CI layout (PYTHONPATH=src): 1037 passed, 7 skipped.
• Schema check passes.

axiom: n/a: receipt provenance infrastructure, no policy encoded.

🤖 Generated with Claude Code

[MaxGhenis]

Merged at reviewed head 842b71d with a merge commit. Gates: gh pr checks exit 0 (13/13 pass), mergeable, not draft, no change requests. Independent review: round 8 APPROVE at 842b71d (rounds 1–7 findings fixed; the round-7 store regression has three committed tests that fail without the fix). Remaining low (tests for a02/a05/a06/r02/g04 on contrived input) is queued as a follow-up.

🤖 Generated with Claude Code