test(prompt-ontology): ADR-022 fingerprint/scope/surface/capability harness (#1726)

[ozand]

Summary

Adds tests/test_prompt_ontology.py, the ADR-022 rule-4/5 test harness named in issue #1726: fingerprint-once, AGENTS.md scope, surface parity, and no-guidance-without-capability, all built over the assembled loop-profile system prompt (release root at IDENTITY.md/SOUL.md/goals.md/USER.md/OPERATING.md + a fixture instance workspace).

What passes today

• Fingerprint once (system-block half): each of skip / mutation surface / branch / runner / identity / budget / final-JSON matches exactly one system block. Regexes were proven against the real release files, not copied verbatim from the issue's draft table — the issue's never (checkout|push) does not match OPERATING.md's real branch sentence, and You are the does not match IDENTITY.md's real self-declaration (it would also false-positive on SOUL.md's "You are an honest instrument"). Both were widened; see the module docstring for the exact greps.
• A deliberate-duplicate fixture (skip rule restated in AGENTS.md) fails, naming both operating and agents and the count.
• AGENTS.md scope: mutation_policy.agents_md_scope_violations() against a compliant target-shape fixture, an over-length fixture, and a runtime-heading fixture.
• Surface parity: OPERATING.md's ## Mutation surface section against MUTATION_POLICY.render_bridge_surface_block() (reusing tests/test_operating_md.py's _section helper rather than duplicating it), a mutated copy fails with a diff, and Allowed targets: is confirmed absent from every other release-root *.md file and every other assembled-prompt block.
• No guidance without capability: every tool name in the assembled prompt's prose is in EXECUTOR_TOOL_NAMES, excluding the loop's own suppressed builtins (mirroring bridge.py's local _LOOP_EXCLUDED_SKILLS = ["weather", "tmux", "clawhub", "cron", "summarize", "github"] — without excluding these the fixture skills catalogue includes cron, a false positive, not a real one).

xfail(strict) — three cases, three reasons

1. test_fingerprint_user_message_only_budget_and_task — #1723 part (b) pending. build_task's user message still carries its own skip/surface/branch/runner/final-JSON literals; PR fix(bridge): remove build_task cycle-rule literals now that OPERATING.md is loaded (#1723 part b) #1742 is already open for this (fix(bridge): remove build_task cycle-rule literals now that OPERATING.md is loaded). Once it merges, this test flips to a hard failure until the marker is removed.
2. test_agents_md_scope_violations_against_real_instance_fixture — #1730 part 1 pending. The real instance AGENTS.md fixture (tests/fixtures/instance_agents_md/AGENTS.md, refreshed via gh api repos/ozand/eeebot-self-evolving/contents/AGENTS.md --jq .content | base64 -d) is 192 lines with 8 runtime headings today, matching the issue's stated baseline exactly.
3. test_search_memory_status_contract_appears_only_in_operating_tools_block — new finding, no issue filed yet. Proving the capability assertion surfaced that the search_memory status contract ("complete/partial/unavailable") is restated in two places inside the assembled prompt: OPERATING.md's ## Tools section and nanobot/agent/memory.py's MemoryStore.MEMORY_SEARCH_POINTER (folded into the loop's memory block whenever memory/index.md is non-trivial). This is a real ADR-022 rule 5 violation, but fixing it means editing the protected OPERATING.md or the shared nanobot/agent/memory.py — both out of this issue's "do not touch" scope — so it's xfail(strict) with the mechanism named in the reason, flagged here for a follow-up issue.

Constraints honored

• Did not modify nanobot/agent/context.py, nanobot/runtime/bridge.py, OPERATING.md, or any other release file.
• tests/fixtures/instance_agents_md/AGENTS.md and memory_index.md (already present untracked from the prior dead attempt) verified byte-identical in content to a fresh gh api pull before committing — no refresh needed.
• ADR-022's Test Contract section gets a one-line status note.
• Ran only python -m pytest tests/test_prompt_ontology.py tests/test_operating_md.py tests/test_ontology_loader.py (Python 3.13, since 3.9/3.11/3.12 on this host lack loguru/pytest) — 26 passed, 3 xfailed, no regressions in the other two files.

Test plan

• python -m pytest tests/test_prompt_ontology.py tests/test_operating_md.py tests/test_ontology_loader.py — 26 passed, 3 xfailed

[ozand]

Rebased onto 023a38f7 (#1742, #1723 part b) and fixed two things the rebase exposed:

1. The user-message fingerprint assertion was xfail(strict=True, reason="#1723 part b pending") and it XPASSed once the literals were gone — the marker did exactly what it was for, so it is dropped rather than left to fail as an unexpected pass.
2. The capability assertion went red on the word message. fix(bridge): remove build_task cycle-rule literals now that OPERATING.md is loaded (#1723 part b) #1742's iteration-budget pointer (added in review, ae10173c) said the number is stated in "your task message", and message is the name of a tool the executor was never given — the assertion is right to reject it. The sentence now points at the task section and names the line instead, and message appears nowhere in OPERATING.md.

38 passed, 2 xfailed locally (tests/test_prompt_ontology.py tests/test_operating_md.py tests/test_ontology_loader.py tests/test_task_prompt_hygiene.py). The two remaining strict xfails are #1730 part 1 (instance AGENTS.md is 192 lines) and the search_memory contract duplication you found — filing that one as its own issue now.