corpus-report scores `nix develop -c bash` as an interpreter hand-roll, so BUILD NEXT names the sanctioned path

[thedavidmeister]

Run against the current producer corpus, corpus-report recommends:

BUILD NEXT: interpreters

with the series behind it:

metric	first	peak	latest	runs	shape
interpreters	19	28	28	4	holding
probes	5	6	4	4	falling
raw gh api	1	1	0	1	quiet
tarball / harness scaffold / dispatcher source	0	0	0	0	absent

The count is the prescribed pattern

Across the 4 traces that have tool calls (20260816T170156Z, 20260817T000148Z, 20260817T050103Z, 20260817T110404Z):

total Bash calls                      1293
nix develop … -c bash                   49
any -c payload naming an interpreter    49

The two are the same 49. CORPUS_INTERPRETERS includes bash and sh, and interpreter_runs deliberately resolves wrappers so bash --norc -c is caught — which means nix develop .#x -c bash -c "…" scores as reaching for an interpreter rather than for a tool.

That shape is what the prompt INSTRUCTS:

forge/cargo/node via nix develop

Command-position interpreters across the same 4 traces are 5 python3 and 4 bash — and at least one of those python3 strings is a call block-python.sh REFUSED, recorded in the trace as an attempt (blocked: Python … This org’s tooling is Rust), not as an invocation.

So the metric’s headline number is ~49 sanctioned toolchain accesses plus a handful of interpreter reaches, most of which the guard already stopped. Acting on BUILD NEXT: interpreters means removing the prescribed way to reach the flake.

Why this matters beyond one wrong line

BUILD NEXT exists so a reader does not have to re-derive the ranking. A recommendation that names the sanctioned path spends a session’s attention on a non-problem, and the corpus’s real message — that probes are falling and tarball / scaffold / dispatcher are absent, i.e. the hand-rolls previous readings targeted are gone — is buried under it.

Suggested fix

Count what the metric is named for: a run reaching for an interpreter INSTEAD of a tool.

• Exclude an interpreter that is the -c payload of a nix develop / nix shell invocation. That is toolchain access, and it is what the prompt asks for.
• Consider whether bash/sh belong in the list at all, given a compound shell command is how the harness invokes everything. python/node/ruby/perl/deno are the reaches the issue behind this metric was about.
• A blocked attempt should not count as a hand-roll the tooling failed to absorb — the guard absorbed it. If attempts are worth counting they should be reported as a separate column (refusals are their own cost: the denial sits in context and is re-read every turn), not folded into the same number.

Done when

• BUILD NEXT does not name a pattern the prompt prescribes.
• The interpreter count distinguishes toolchain access, a genuine interpreter reach, and a reach the guard refused.
• A test pins each of those three against a fixture trace, since all three are one number today and nothing would show them diverging.