2026-08-19 - The diff loop, which is what the format was for

Done. Built the capability the report format has been paying for since D003.

• tools/diff.py - compares two reports, exits non-zero on regression.
• make diff BASELINE=... - change something, re-run, ask whether it helped.
• New build record, stamped at compile time (make BUILD_ID=...), so a diff can separate "the probe changed" from "the platform changed".
• tools/test-tools.sh - pins the exit codes of both tools, in make check and CI.
• CI stamps the commit into every report.

Verified. 8/8 tooling tests. Diffed a real pair of runs - the libm and no-libm host builds - and it reported exactly the truth: 7 improved, 0 regressed, 70 unchanged, 20 symbols newly present, tally: pass +7, skip -7. Reversed, it reported 7 regressions and exited 1. Against itself, silence.

Surprises.

• The format had been justified by a tool that did not exist. Every awkward choice on the C side - stable identifiers, one record per line, raw numbers instead of prose

  • was argued for on the grounds that an agent would diff runs. Nothing could diff runs. The cost was being paid and the benefit was theoretical.

• Defining "regression" was the whole design. The obvious reading - a check that is failing - is useless here, because under an early emulator almost everything fails and would report as a permanent regression. It has to mean got worse, which makes a run where everything fails and nothing changed a clean exit. Correct, and it looks wrong until you think about when the tool is used.

• skip ranks below fail, not above it. A check that stopped running tells you less than one that ran and failed. Without that ordering, losing coverage reads as neutral - and deleting an inconvenient check would read as an improvement.

• Testing the tools against a degraded copy of a real report, not a fixture. A fixture drifts from the format while its tests keep passing, which is precisely the failure being guarded against.

• PowerShell ate $? before bash saw it, so two rounds of "exit code" output were reporting PowerShell truth values rather than the tools. The exit codes were only actually verified once a script inside the VM asserted them - which is the reason test-tools.sh exists rather than a series of ad-hoc invocations.

Not done. The census is still my recollection of what matters rather than a real title import table. Struct layouts, NID verification and running under an emulator all remain blocked.