Add pdr diff

[ILIASTEKEOGLOU]

Summary

qstriage pdr diff BEFORE AFTER [--format markdown|json] [--output FILE] [--overwrite]

• Both documents are verified first (0.2 and 0.3 rules). If either fails,
  the command prints which one and exits 1 without a comparison.
• Records are matched by record_id; decision and observed_state fields
  are compared.
• Provenance items compared: input file (source_hash), policy pack
  (policy_pack_hash), algorithm registry (registry_hash), QSTriage
  version (engine.version), PDR format (pdr_version). When more than one
  differs, the report states that a decision change cannot be attributed to
  one of them. No cause is inferred.
• Between different pdr_version values, fields present in only one
  document are listed as format differences, not as changes.

Report layout (Markdown, default)

1. Result: decisions changed, records added and removed, verification of
   both files.
2. What changed between the two runs, in plain words.
3. Decisions that changed: action, algorithm status, human review, reason
   codes, per asset.
4. Records whose observed state changed without a decision change, listed
   separately.
5. Format differences (not decision changes).
6. Field details table.

The report shows what the records contain and does not recommend actions.
Text taken from the documents is escaped with the same helpers as the
Markdown report. --format json carries the same content with document,
policy-pack and registry hashes.

Exit codes

0: comparison produced (the JSON field changed states whether anything
changed). 1: read, verification or write failure. 2: invalid option.

Verification

• 13 tests: identical documents, single-cause and multiple-cause
  attribution, 0.2 to 0.3 and 0.3 to 0.2 format differences, added and
  removed records, tampered document refused, duplicate record_id
  refused, Markdown escaping of untrusted names, JSON content, CLI exit
  codes, no-clobber output, inputs unchanged.
• Local runs: Linux Python 3.11 and 3.13, 1454 passed, 3 skipped; Windows
  Python 3.11, 1453 passed, 4 skipped. Comparing the stored PDR 0.2 fixture
  with a PDR 0.3 in which one algorithm became X25519MLKEM768 reports 2 of
  5 decisions changed and exits 0.

Documentation

docs/usage.md, docs/pdr-contract.md (Comparison), input limits in
docs/input-contracts.md and the README, CHANGELOG (Unreleased, Added).