Docs in sync with release-v0.2: prune removed features, current names

[ThrudPrimrose]

Synchronizes the docs on release-v0.2 with the code that ships. Docs, comments and one YAML header only; no code behaviour changes. Merged with origin/release-v0.2 (incl. "Prune checks": no remaining mention of check_manifest_structure.py, check_large_files.py, check_reference_naming.py, run_smoke.sh or enroot).

Removed (text describing features that are gone)

• docs/configuration.md: the duplicate "Benchmark settings (config.yaml)" section (config.coerce, config.settings(), config.get(key, default), "keys are not validated", "reader's default"). One section remains.
• docs/runtime.md: the empty runtime.backend -> podman default and containers.DEFAULT_BACKEND (the key is a Literal now).
• docs/extending/protocol.md: control_setups in studies.yaml.
• docs/plotting.md: the DEFAULT_BASELINE fallback (population.one_denominator refuses a slice with no denominator).
• docs/token_accounting.md: the episodes.MIN_TOKEN_FOLD gate.
• docs/prompts.md: prompt_hash, the prompts table and __<variant> filenames.
• docs/requests_api.md: "old envelope fields refused naming the new spelling" (Submission refuses every undeclared field).
• CONTRIBUTING.md: the manifest-structure hook in the hook list; "or hint" (hints are gone).
• hpcagent_bench/config.yaml comments: "" = podman.

Renamed (stale names -> current)

• docs/translator_desugarings_and_tool_bugs.md: frontend._inline_module_constants, lowering._EnumerateZipRewriter, _MethodCallRewriter, _lp_normalize_index_access, _fix_real_scalar_dtypes, lib_nodes._retarget_scalar_accumulator, numpy_desugar._OuterBroadcastPeel, lib_nodes (_remap) -> the public names (lib_nodes.pad edge_clamp).
• hpcagent_bench/docs/library_requests.md: sandbox.DISTRIBUTED_CONTRACT_LIBRARIES -> grading.distributed_libraries / sandbox.distributed_contract_libraries.
• docs/measurement_statistics.md: seeds.secret_first -> hidden_seeds.secret_seed_first.
• helpers/skills/python-to-numpy/SKILL.md: UNGATED_SUBTRACKS -> UNGATED_TAGS.
• hpcagent_bench/harness/hidden_tests/README.md: context.py (gone).
• CONTRIBUTING.md: prompts/sections/lang/ -> prompts/lang/.
• docs/extending/packets.md: example class CpfTool -> CpfToolCpu; docs/writing_an_agent.md, docs/prompts.md: the CPF tool/source packets by their device keys.
• hpcagent_bench/config.yaml comments: envs/display_names.yaml -> @llm keys of models.py; task.host_only_setup -> task.setup_declared_host_only.

Rewritten (current behaviour, briefly)

• docs/configuration.md "Benchmark settings": config.yaml is the only place a default is written, validated by the pydantic models of config.py; resolution order; env parsing by type (config.parse_env); unknown keys/env names are errors. config.yaml's header comment says the same.
• docs/requests_api.md action table: which operations lower with startup.grid (distribute, replicate, split/block/permute and inverses, place) and which run on one node only (pad, shuffle, linear_shuffle, zip, unzip), instead of "parsed, refused".
• docs/extending/benchmark.md initializer: rng is a CounterGenerator (not a numpy.random.Generator); a GPU grade builds its inputs on the device with xp = cupy, bit-identical to the host, so elementary functions and floating reductions come from support/exact_math.py.
• docs/extending/packets.md: the four CPF keys (cpf-cpu-src, cpf-gpu-src, cpf-tool-cpu, cpf-tool-gpu) stated once; cpf-src / cpf-tool are the device-neutral aliases packets.by_device resolves.
• docs/hf_dataset_and_harbor.md: the HF export runs in the coverage job (there is no hf-export step).
• CONTRIBUTING.md, experiments/LAUNCH.md: pytest -m smoke (tests/test_smoke.py) is the pre-launch check on the login node.
• docs/extending/skills-and-tools.md: applies.explicit, and the frontmatter is the pydantic packets.SkillFrontmatter / Applies (unknown keys refused). CONTRIBUTING.md: a compilers.yaml block is a languages.CompilerBlock.

No doc file was removed: none covers only removed features.

Checks

• sphinx-build -W --keep-going -b html docs passes; helpers/scripts/checks/check_doc_links.py passes.
• pre-commit run --files <changed> green.
• Doc-pinning tests (test_docs_match_code, test_protocol_samples, test_skill_content, test_check_no_corpus_leak, test_skill_triggers, test_skill_packet_leaks, test_skill_applies, test_prompt_skills, test_experiments, test_config, test_cluster_prompt_sources, test_check_repo_rules, test_packets, test_vocabulary) pass. test_cpf_pages_vocabulary fails locally only because the local venv's dace is not the f97160b pin (SEQUENTIAL_ON_CPU missing); the CPF pages were not touched.

Left as is (needs a decision)

• hpcagent_bench/support/update_names.py and its doc lines (docs/extending/optimizer.md, columns.py) still ship: the framework/canon DB names stamp uses it. Not pruned.
• docs/code_reuse.md is a consolidation audit pinned to old commits (proposal rows, line numbers at 934c5b98d); kept, not refreshed.
• docs/observations.md blank meanings "recorded before the stamp/column" describe rows a fresh hb-results-v1.0 DB may no longer produce; kept.

🤖 Generated with Claude Code