Paraphrase owner quotations in the new-WSL layer-consensus record and regenerate the manifest and handbook - #1159

The GPT read of #1128 at bd3eecf (blocker 1, P1) found owner quotation
spans in the generated handbook's agent-messaging record. They came from
the layer-consensus record: assemble_manifest.py copies its rows and
amendments into the definitive manifest, and build_new_wsl_handbook.py
copies each manifest slot whole into the handbook, so the fix is in the
canonical input and the outputs are regenerated from it.

Layer-consensus record (consensus.json): six owner quotation spans in
five fields become dated, attributed paraphrases of the form: the owner,
<date> (paraphrased): ..., with the date taken from each record's own
context:
- authorization: verbatim -> paraphrase (2026-10-02)
- amend_rows[4] agent-messaging amendment: owner_verbatim ->
  owner_paraphrase (2026-10-02); its decision, text, arms and experiment
  contract are unchanged
- wave2 context-supply interim authority: verbatim -> paraphrase
  (2026-10-03)
- wave5 agent-messaging owner default: two spans in its reason
  (2026-10-02 and 2026-10-03) and one in its posture (2026-10-03); the
  public record of the answer stays linked by URL

Regenerated through the existing generators:
- assemble_manifest.py (write; --check exit 0): definitive-manifest.json
- render_tables.py --check: the 2026-10-01 record's tables stay current
- build_new_wsl_handbook.py --write, then --check exit 0:
  docs/new-wsl-handbook.json and docs/new-wsl-handbook.md
The handbook loses five owner spans: the four the read named (base lines
851, 933, 934) and one in the context-supply record's overturned interim
(base line 2383), which the read did not list.

Bindings: the handbook receipt's current outputs, a regenerations entry
and a previous_outputs entry; manifests/evidence.json re-registers the
eight changed hashed files; scripts/validate.py passes.

Regression: tests/test_new_wsl_handbook.py OwnerQuotationTests checks by
structure, with invented prose only: a field that holds the owner's
words, or a sentence that attributes a double-quoted span to the owner
or the user, fails; a dated paraphrase and separately defined protocol
or contract text pass. Against the base tree it fails with 13 locators
(consensus record 5, manifest 4, handbook 4); after the change all three
tests pass.

tests/test_new_wsl_definitive_defaults.py reads the paraphrase fields,
checks the consensus decision record with the same attribution rule, and
its wave-3 relay check no longer holds any of the owner's words (it
checks the record's block quote instead). assemble_manifest.py: comment
only, the interim authority now carries a dated paraphrase.

Not changed: the owner batches' authority.verbatim fields (wave3, wave4,
wave5) stay in consensus.json. assemble_manifest.py requires them
(OWNER_AUTHORITY_FIELDS, line 389) and refuses a batch whose hashed
relaying decision record does not contain them (check_owner_batch,
lines 542-544); they
never reach the manifest or the handbook. Changing that relay contract
needs its own ruling. The regression keeps exactly those paths, and only
while the assembler requires the field.

This supersedes, for the canonical inputs, the integration resolution
that chose exact quotation (evidence/artifacts/
final-architecture-round2-20261004/integration-resolutions.json:543,
:630, :801). No history rewrite: earlier commits keep the old text.

Co-Authored-By: Claude Opus 5.5 <[email protected]>