consumer-py — the second-language consumer (X1 milestone M5)

Python 3, stdlib only (hashlib, json, re, pathlib) — the fixtures lab's
zero-dep rule, in a second language. Run from this directory:

python consume.py   # re-run the frozen suite     -> vectors: 39/39 pass
python test.py      # the P1–P6 property suite    -> consumer-py: 6/6 pass

What this proves

Two implementations — ../fixtures/consume.mjs (JavaScript) and this
consume.py (Python) — share no code. They share only the data: the 39
vectors under ../fixtures/vectors/, the manifest, and the reason-register
vocabulary. Both consumers load the same vectors, re-derive every digest,
re-run the constructions, and reach the same outcome and the same reason
code, byte for byte, on every vector (P1), including the same per-family
counts (PR-CTX 2, PR-EPO 3, PR-FLB 2, PR-HLD 1, PR-LIV 5, PR-PER 1, PR-SHW 8,
PR-TGF 7, PR-UNQ 10 = 39).

That is the X1 interop claim demonstrated, and §26's requirement instantiated:
independent implementations instantiate the decisions consistently. The
fixtures — not the JavaScript — carry the decisions. X1's M5 milestone
("a second implementation in another language would rewrite exactly this file
and nothing else" — consume.mjs header) is discharged by this directory.

Suite size note: the vector count is 39 (9 families), exactly as
manifest.json pins (vectorCount: 39); the 40th .json file in the vectors
tree is the manifest itself.

Canonical-JSON byte-compatibility — what actually mattered

P3 is the critical property: all 29 embedded digests (descriptor digests
bound into transcripts, transcript digests, the show transcript digest, the
six member commitments of the accept show, and the two card:<digest[:12]>
versions in the PR-CTX migration record) re-derive byte-identically from the
Python canonical encoding. What had to match the JS canonicalize():

• Separators: exactly {"k":v,"k2":v2} and [a,b] — no spaces anywhere.
  (Python's json.dumps default of ', ' / ': ' would break every digest;
  the encoder here builds the string by hand like the JS does.)
• Key sorting: JS Object.keys(...).sort() sorts UTF-16 code units;
  Python sorted() sorts code points. Identical for ASCII keys — and every
  key in the suite is ASCII. A key containing astral-plane characters would
  diverge; recorded here as the known boundary of the port.
• String escaping: JSON.stringify escapes " and \, uses the short
  escapes \b \t \n \f \r for those controls, \u00XX for other chars
  < 0x20, and leaves non-ASCII raw. json.dumps(s, ensure_ascii=False)
  produces exactly that. (With the default ensure_ascii=True, any non-ASCII
  byte would diverge — the suite's canonicalized objects happen to be pure
  ASCII, but the port does not rely on that.)
• Numbers: all numbers in the canonicalized data are integers (epoch
  seconds); JS prints them without a decimal point and so does Python for
  int. The port also renders integral floats the JS way (1 not 1.0)
  and rejects non-finite numbers, but no vector exercises a float — untested
  path, flagged honestly.
• undefined filtering: the JS drops undefined-valued keys. JSON-loaded
  Python dicts cannot contain such a value, so no code was needed — noted so a
  future non-JSON caller does not assume the filter exists.
• Hash discipline: runtime 01's H = SHA-256 over parts, each prefixed
  with its 4-byte big-endian byte length (so H(a,b) != H(a||b)), parts UTF-8
  encoded; domain-tag strings copied byte-exact from the sources.

The register port (reasons.py)

The register is treated as data: the 66 exact codes and 29 family prefixes
are parsed out of ../fixtures/reasons.mjs as text (never imported or
executed), and the vectors' own expect.reason codes (22 distinct across the
suite) are collected alongside. Only the mechanism is re-implemented: exact
stratum + family stratum (isKnownReason = exact match, or family prefix +
non-empty parameter). P2 confirms every non-accept reason in the suite is
known to the port. Parsed counts (66 + 29 = 95, version v2) agree with the
JS suite's own F10 result.

What was NOT ported (nothing any vector exercises)

• rt01: nothing omitted — enrol / nullifier / ContextRegistry all used.
• rt07: encounter(), Mage.propose(), dreamCycleTurn(),
  TrustGraph.neighbours/connected — vectors carry the candidate and the
  shared compression pre-computed; only joinCommunity, rdid,
  Swordsman.prove, roster, and edge add/has are needed.
• show-composition: makeShow (and so bundle-profile-mismatch),
  matchBundle, completionEvidence (taskcontext-not-outcome-evidence),
  jointDisclosureRecord, STATEMENTS/OBSERVERS content — verifyShow
  never consults them for a verdict. The transcript-invalid branch of
  verifyShow IS ported (it is reachable), but no vector triggers it.
• context-card: the provingMode q5 branch and custom-justification
  merging are ported but unexercised; checkLegibility's failure families
  (unrendered-field: etc.) can fire but no vector makes them fire.
• canonical: transcripts with delegationRef (no PR-DEL vector).
• Register-wide: the 29 family-prefixed codes and the v2 instrument codes
  (quiet-presentation, rotation, erosion, mediator, multi-issuer, guardian)
  exist only as vocabulary here — no vector exercises those constructions,
  so no Python port of those six labs was attempted (that is the JS suite's
  F10 live-trigger job, out of M5's scope).

Divergences found between the JS and the data

None. Every embedded digest re-derives (P3: 29/29), every outcome and
reason matches (P1: 39/39), the manifest agrees with the tree, and the lint
regex ports reproduce all 7 prohibited hits with the frozen reason codes while
passing all 3 narrow-language samples (P5) — including the unless-guarded
collusion pattern. The JS suite (node ../fixtures/test.mjs) still passes
13/13 untouched. One near-miss worth recording: JS \s and Python \s differ
slightly at the Unicode margins (JS includes \v and some exotic spaces in
both engines — they agree on everything the lint texts contain), so the lint
port's equivalence is demonstrated on the frozen samples, not proven in
general.

Pointer

X1 exploration doc:
X1 — Conformance Fixtures —
milestone M5, "second-language consumer": this directory discharges it.
Counterpart consumer:
../fixtures/consume.mjs; vector producer: ../fixtures/extract.mjs;
register: ../fixtures/reasons.mjs (v2); suite: ../fixtures/test.mjs
(13/13, unchanged by this build).