fix(docs): correct entity-resolution framing around merged provenance (#427)

[jasonssdev]

Closes #427.

Unblocks #379 criterion 1, the P0 validation gate's only remaining criterion that needed code.

What this is

#427's product decision was already made: lean on entity resolution, do not build a cross-source synthesis step. This ships the two halves that decision implies.

Half A — the accumulation path was already sound

The 8-source run's evidence said the union path was broken. It is not. Verified with file:line evidence:

• provenance is not in _SPECIAL_KEYS (src/openkos/model/okf.py:1029-1035), so it falls to the list branch and is unioned via _union_dedup (:1040-1043).
• Both measured pairs were nominated correctly: "MCP Workflows"/"Model Context Protocol" by the ACRONYM tier, "Skill Creation Process"/"Skill Creator" by the LOW tier (SequenceMatcher("creator","creation").ratio() ≈ 0.8 clears the 0.75 threshold).
• Both were adjudicated DIFFERENT correctly — adjudication.py:52-58 instructs the model to reject part-whole and aspect relationships as duplicates.
• _MAX_CANDIDATE_GROUPS = 50 was never the binding constraint; the run produced 2 groups.

Candidate generation worked, the adjudicator was right, and the corpus simply contained no true duplicate. So Half A is a proof obligation, not a repair — which is what #379 criterion 1 anticipated: "If no natural SAME pair arises in the corpus, construct one."

The new test constructs one. It runs the real find_candidates → adjudicate_candidates → merge chain, stubbing only OllamaClient — unlike its neighbours in the same module, which monkeypatch the resolution functions themselves.

The test was mutation-checked

It passed on its first run with no production change. That is expected for a regression guard over existing behavior, but it is also the signature of a test that asserts nothing. So a mutant was injected at okf.py:1043, replacing the union with list(survivor_list):

AssertionError: assert ['sources/not...-enchiridion'] == ['sources/not...l-with-maria']
  Right contains one more item: 'sources/call-with-maria'
1 failed, 93 deselected

The guard detects a real regression. The mutant was reverted and the tree restored byte-for-byte.

Half B — three prose corrections, zero fixture changes

A first analysis concluded the good-life-demo fixtures' two-source provenance: was unreachable and should be trimmed to one source. That was wrong on two counts, and the exploration records the correction:

1. stoicism.md's body carries [1]/[2] markers and a Citations block naming both sources. Trimming would leave the document citing a source its own provenance denies.
2. The two-source list is reachable — two sources each yielding a Concept titled "Stoicism" share one normalize_key(title), which is the HIGH tier's trigger; a SAME verdict merges them and unions the provenance. It is a picture of entity resolution having fired.

What was actually wrong is the framing around the fixtures:

File	Problem
examples/README.md:5	"what the MVP 1 ingest should produce" — ingest alone writes single-element provenance literals (cli/main.py:1955, :2795)
docs/knowledge-object-model.md:91	quoted verbatim with no indication the two-source provenance is the product of a merge
docs/knowledge-object-model.md:311	"as new sources arrive, the engine rewrites existing objects" — overclaims passive, automatic behavior

The fixtures themselves are untouched.

Left out, deliberately

The one genuine reachability gap found is strict okf_type partitioning in _keyed_docs_by_type (candidates.py:208-246, pinned by test_cross_type_identical_normalized_title_produces_no_candidate): a duplicate whose two sources were classified into different OKF types is never nominated. It is filed separately as issue #437 rather than folded in here — it breaks CandidateGroup's single-okf_type contract and changes a partitioning invariant two other call sites depend on.

Verification

pytest -q              -> 3706 passed in 120.98s
ruff check .           -> All checks passed!
ruff format --check .  -> 171 files already formatted
mypy . --strict        -> Success: no issues found in 171 source files

Review lineage review-bc49ea0e43232da2 — review-reliability lens, inspection completed, zero findings; pre-commit/pre-push/pre-pr gates all allow.