feat(ENV-AGNOSTIC-W1-TOOLING): resolve or refuse in scripts/ and tools/, and keep the 9 recorded hosts literal

[localai-bot]

Wave 1 of #1190, the campaign scoped by env-agnostic.md. It applies that row's classification rule to scripts/, tools/ and third_party/README.md, and it lands .agents/specs/env-agnostic-w1-tooling.md first so the classification is on record before any file moves.

Re-derived at this row's own base rather than quoted, the campaign's query returns 19 files and 22 hits over those paths. 13 are configuration and become ${KEY}; 9 are provenance and stay exactly as they are. The same query now returns 9, and every one is a dated measurement or a named-instrument argument. The split is not by file: three near-tie-gap scripts hold a measured sentence at :47 that stays and a copyable oracle-venv command at :22 that converts, and the query matched only the first.

Five entry points answered instead of failing

A wrong default is not an inconvenience here, it is a false green.

tools/parity/dump_{gdn,moe,qwen3_5_mtp}.py defaulted --pin to one operator's checkout. That flag names the tree the ORACLE MATH executes from, so on another box it resolves to something else and the dump writes the wrong provenance into the manifest, which no goldens comparison can see. They now take ${VLLM_SOURCE}, print which tree they resolved, and refuse by name with exit 3 before importing anything.

scripts/regen-triton-aot.sh still carried the ~/cutlass_probe default that scripts/dgx-bringup.sh had just lost, and .agents/environment.md:388-400 measures a configure that misses CUTLASS as moving the SACRED test_qwen27_paged_engine from 235/235 to 234/235 with the source untouched. This one also REGENERATES artifacts that get committed, so the wrong toolchain lands in the tree. It now sources .env with the shell winning over the file, resolves the interpreter from ${VLLM_ORACLE} in either shape .env.example declares, takes ${DEVICE_TOOLKIT_ROOT} instead of prepending one CUDA prefix unconditionally, and refuses VLLM_ORACLE and CUTLASS_DIR by name.

scripts/laguna_longctx_bench.sh defaulted its checkpoint and its binary into one $HOME layout, so another developer's run took the GPU lock and dropped the page cache before finding nothing. Both operands are now required.

The query cannot see the defect it was written to find

scripts/regen-triton-aot.sh matches none of the campaign's four patterns. It is in this wave because the campaign row wrote the debt into the file's own text at :20-23, naming this row. A hard-coded default names a PATH and the query names a HOST.

Widening the query over the same two directories moves 19 files and 22 hits to 73 and 174, and 12 of those lines in six shell scripts are the exact ${KEY:-<one operator's path>} shape the worked example removed. The same blind spot covers W3 and W4, so the campaign's claim that its five waves partition 227 files is a claim about the query's 227 rather than about the tree. This row converts the one file it owns by name and leaves the rest, because two of them argue instrument identity the way scripts/mtp-k-gt-1-neartie-gap.py does and must not be swept blind. Filed as #1308 and owned under the spec's ## Owed.

The half that guards the records

tests/scripts/test_env_agnostic_tooling.py is 11 cases and needs no torch, no cmake and no GPU: the dumpers run under stub modules and every refusal is asserted before the step that would need hardware.

Its second half pins the 9 provenance sentences by content. A blind sed of the tree is the cheapest way to close #1190 and the only way to falsify a recorded measurement while doing it, and until now nothing anywhere would have gone red. Every other assertion in the file would pass such a sweep.

Evidence

Red before green at base 5c8671c50, suite written and no file converted: Ran 11 tests, FAILED (failures=58), ten of eleven cases red. The one that passed is the provenance guard, which is correct — it protects what must not move. Green after: 11 tests OK.

Four mutations on the committed tree, each restored and confirmed byte-for-byte by git write-tree returning e2d6c00e5 again with git status --porcelain empty, and each compiling so that none is a build failure reading as a pass:

Mutation	compile_rc	Failing cases
sed 's/dgx\.casa/${GATE_HOST}/g' over the six provenance files	0	6, all ProvenanceTests
reintroduce ${CUTLASS_DIR:-${HOME}/cutlass_probe} in regen-triton-aot.sh	0	3
restore --pin default /home/mudler/_git/vllm in dump_gdn.py	0	3
delete the resolve_pinned_source call site in dump_moe.py	0	4

The last is the reachability mutation: resolving the literal inline keeps parity_env.py compiling and its behaviour intact, and the suite still goes red, so the gate measures the entry point rather than the helper.

No hardware gate applies. This row moves where a value comes from and changes no kernel, no dtype, no allocation and no token, so it claims no CUDA, SACRED or throughput gate.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5 [claude-code]