Remove dead pointers and sibling-repo references left by the public move

[deepfates]

Two kinds of residue from the cleanup PRs: paths that name files no longer in the tree, and documentation that routes a public reader to the owner's private repositories.

Stale pointers

• benchmarks/authorities.json generated_from: dropped benchmarks/claims.json, docs/internal/UPSTREAM_SURFACE_MAP.md, docs/internal/COVERAGE_MATRIX.md and docs/internal/PARITY_VALIDATION_PROGRAM.md, which are gone; repointed the two that survived the move to docs/differentials/INSTRUCTION_OPTIMIZER_FIDELITY.md and docs/differentials/RESEARCH_LANDSCAPE.md.
• benchmarks/authorities.json dataset_protocol.references: removed the seven docs/internal/PARITY_VALIDATION_PROGRAM.md#... references; where that emptied the list the status moves from protocol_defined/pinned to gap, because nothing in the repository defines the protocol any more (families optimizer_few_shot_search, optimizer_instruction_search, optimizer_simba, optimizer_gepa, runtime_performance); tools_agents_react and optimizer_miprov2 keep their statuses on their remaining references.
• benchmarks/authorities.json local_differential.artifacts: removed fourteen entries under benchmarks/results/ and benchmarks/evidence/, none of which was ever tracked; six families whose only evidence was such an entry move to status: gap, artifacts: [], and three notes that still spoke of an admitted artifact now say the artifact is unpublished history.
• docs/differentials/COMBEE_FIDELITY.md, CONFIDENCE_CALIBRATION.md (two), MULTIMODAL_FIDELITY.md, PLAYBOOK_OPTIMIZER.md and RLM_FIDELITY.md (two) cited result JSONs under benchmarks/results/ as retained evidence; each sentence now states the observation with its date and says the result JSON is unpublished history not retained here.
• examples/matched_instruction_optimizers_trec/README.md: the recompute command now matches docs/BENCHMARKS.md exactly and reads the committed rows under examples/matched_instruction_optimizers_trec/data/; the sealed campaign artifact is stated as unpublished history rather than given a deleted benchmarks/evidence/archive/... path.
• benchmarks/config/support-ticket-lift-openrouter-free-v2.json, -v3.json, local-mlx-lifecycle-banking77-v2.json and -v2-deployment-continuation.json keep their output_artifact and predecessor.artifact strings, which are a write target and a preregistration record respectively, and each gains one note key saying that benchmarks/results/ is untracked and a predecessor named there is unpublished history.

Self-containment

• README.md "Where this fits" is now four sentences about Imp alone: a library, its deepfates/ex_mcp fork dependency as the one MCP and ACP implementation it uses, no protocol endpoint on ordinary startup, and a host application owning product lifetimes.
• decisions.md: dropped the "(workshop)" tags, restated "Workshop AGENTS.md" as "Owner ruling, 2026-09-11" (that file does not carry these rulings, so citing it would have been false) or as CONTRIBUTING.md and .github/workflows/ci.yml where those do; removed the sibling names from the ex_mcp pin row and the "Haven preset" mention; rewrote the "Nothing leaves these repositories" row as outside contributions welcome through GitHub with ex_mcp patches upstreamed only by the owner.
• decisions.md ticket row: it cited a CONTRIBUTING.md sentence and a 57-file .tickets/ directory that are both gone, so it is now recorded as resolved.
• examples/matched_instruction_optimizers_trec/: "workshop spend ceiling" is now "spend ceiling" and WORKSHOP_CEILING is SPEND_CEILING across README.md, contract.exs, run_paired.py and paired_coordinator_test.py. The renamed contract.exs launch-status value and the renamed coordinator output key have no readers.
• lib/imp/mcp/oauth.ex, lib/imp/acp/local.ex, mix.exs, test/mcp_oauth_test.exs, test/react_v2_test.exs: replaced "a Dwell resident, Haven", "a resident application", "Shared constellation reference", "Haven's shape" and "a resident's client" with host-application wording.

Deliberately left

• examples/deployment/data/banking77-mipro-stage1.json embeds 37 repository paths under exposure.scanned_files[].path, some now deleted. It is a leakage-audit receipt; examples/deployment/banking77_mipro.exs reads only condition_id, digests and the three splits, so nothing treats those paths as inputs.
• examples/matched_instruction_family_ifbench/contract-draft.json still has known_workshop_aggregate_before_usd_lte, workshop_aggregate_after_usd_lte and authorized_workshop_ceiling_usd. It is a preregistration record with no readers, and the names describe a spend ceiling rather than any other repository.
• "resident" as in memory-resident (bench/imp/benchmark_truth/hover_gepa_no_merge_plan.ex, test/hover_gepa_no_merge_plan_test.exs, test/local_simba_feedback_trec_example_test.exs), "haven"/"haven't" in the GEPA stopword list, and the "# Haven" markdown heading fixture in test/rlm_interpreter_test.exs.
• benchmarks/config/playbook-live.json was named in the brief but writes to benchmarks/runs/ and benchmarks/checkpoints/, not benchmarks/results/; nothing to change.

Verification, locally, all passing: mix format --check-formatted; MIX_ENV=test mix compile --warnings-as-errors; mix test over the eight named files, 81 tests 0 failures; python3 examples/matched_instruction_optimizers_trec/paired_coordinator_test.py, 6 tests OK; python3 test/python_upstream_authority_registry_test.py, 4 tests OK; MIX_ENV=test mix docs.check, 5 livebooks passed; and the TREC recompute command from docs/BENCHMARKS.md, which agrees with the committed aggregate in full.