Corpus Maintenance

Corpus maintenance

Use this skill when the user asks to maintain, refresh, drift-check, re-pin, or benchmark the authored-source correspondence corpus — the vendored source oracle behind the decompiler's offline benchmark. Produce a short status report: whether the pinned snapshot still corresponds to upstream, whether it needs a re-harvest or a pin bump, and what the benchmark says about the current decompiler.

Run from the repository root. The corpus tooling is the DecompilerHarness (tools/DecompilerHarness), a dev tool, not the shipped dotnet-inspect product; run it with dotnet run --project tools/DecompilerHarness -c Release --.

What the corpus is

The corpus is a vendored JSONL where each row is a real method identity plus a checksum-verified authored member body captured through SourceLink at harvest time. It lets the benchmark grade source correspondence fully offline, so SourceUnavailable becomes a drift signal rather than an expected network outcome. It has two halves:

• CIVIL (Curated Index of Varied IL) — civil/corpus.jsonl, harvested from the 14 pinned assemblies in eng/prepare-decompiler-corpus.sh. This is the stable regression target that grows in lock step with the fixed real-world corpus.
• EVIL (Edge-case Verification of IL Legibility) — evil/corpus.jsonl, an adversarial stress set of the most difficult real methods drawn from a much broader pool (eng/prepare-evil-corpus.sh), difficulty-ranked off the shared IL substrate.

Both live on the vendor/authored-source-corpus orphan branch so the harvested third-party source snapshots never enter main's history. Restore them into a git worktree first:

bash eng/restore-authored-source-corpus.sh

That materializes external/authored-source-corpus. Because it is a worktree on the orphan branch, edits made there (a re-harvest) commit directly to vendor/authored-source-corpus.

The pin, and why it is stable

Every corpus assembly is a pinned published package version, so the only thing that varies across a decompiler change is the tool, never the input. The dotnet-inspect self-assemblies come from the published dotnet-inspect.any package (SELF_VERSION in eng/prepare-decompiler-corpus.sh), not the local artifacts/bin build — that removes corpus drift (#1404) and breaks the circularity where a decompiler change would rebuild both the tool and its own corpus at once.

Two pin axes matter:

• Package version — SELF_VERSION (and the third-party versions in prepare-decompiler-corpus.sh).
• Source commit — each row's SourceLink commit, which is where its authored body was captured.

Pinning buys attribution and reproducibility; "update to latest" is a deliberate, reviewed re-pin (a re-harvest), because pins rot as upstream moves. Drift verification is the bridge that turns "should we re-pin?" into an evidence-driven decision.

Verify drift (start here)

Drift verification re-acquires each vendored row's authored source today and compares it byte-for-byte (newline-normalized) against the stored snapshot. It does not run the decompiler; it audits the corpus, not the tool. Prepare the pinned assemblies, then verify CIVIL, resolving dotnet-inspect's own rows from local git via --repo:

bash eng/prepare-decompiler-corpus.sh /tmp/corpus-assemblies.txt
dotnet run --project tools/DecompilerHarness -c Release -- \
  --verify-authored-corpus external/authored-source-corpus/civil/corpus.jsonl \
  --repo "$(git rev-parse --show-toplevel)" \
  $(cat /tmp/corpus-assemblies.txt)

Each row lands in one of three states:

• Verified — the re-acquired body matches the vendored snapshot.
• Drifted — the body differs; the upstream source, the harvest slice, or the stored row itself has changed. The report names the row with a short line-count/first-diff summary.
• Unavailable — the source could not be re-acquired or sliced (offline with no --repo, commit gone, checksum mismatch, or an extraction regression). The row is surfaced, never silently dropped.

It is report-only by default (exit 0). The run's own integrity still governs the exit code with a named blocker: any corpus row whose assembly was not supplied (unmatchedRows > 0) or a run that evaluated no rows fails, so an empty or partially-unmatched run is never a success. Add --fail-on-drift to make it a fail-closed gate — the run then exits nonzero unless every evaluated row is Verified (any Drifted or Unavailable row fails):

dotnet run --project tools/DecompilerHarness -c Release -- \
  --verify-authored-corpus external/authored-source-corpus/civil/corpus.jsonl \
  --repo "$(git rev-parse --show-toplevel)" --fail-on-drift \
  $(cat /tmp/corpus-assemblies.txt)

Always supply the same pinned assemblies the corpus was harvested from; a missing assembly is a blocker, not a pass.

Interpret and decide

• All Verified — the pinned snapshot still corresponds to upstream. No action; the corpus is a faithful oracle.
• Drifted rows — upstream (or the slice) moved off the pinned commit. This is expected as repositories evolve and is not a bug — the fix is a deliberate re-harvest / re-pin (below), not a code change. Confirm the drift is upstream churn and not a harvester regression before advancing.
• Unavailable rows — usually an outage, a missing PDB, or a bad/incomplete --repo checkout, not corpus rot. Re-run with a complete --repo and network before concluding the source is gone. Persistent Unavailable at a known-good commit points at an extraction regression worth an issue.

Distinguish "the corpus drifted from upstream" (re-pin) from "the harness can no longer resolve known-good source" (investigate the harness).

Re-harvest and re-pin

Advancing the snapshot is a deliberate, reviewed act. To re-harvest CIVIL from the current pins straight into the vendored worktree:

bash eng/prepare-decompiler-corpus.sh /tmp/corpus-assemblies.txt
dotnet run --project tools/DecompilerHarness -c Release -- \
  --harvest-authored-corpus external/authored-source-corpus/civil/corpus.jsonl \
  --harvest-target 12000 \
  --repo "$(git rev-parse --show-toplevel)" \
  $(cat /tmp/corpus-assemblies.txt)

--repo (repeatable) reads each target's authored source from a local git clone, arbitrated by the PDB checksum, falling back to the network on any mismatch or miss. Pointing it at this checkout resolves the dotnet-inspect self-corpus rows with no GitHub round-trip; it also unlocks private, large, or offline source repositories. Third-party rows are checksum-arbitrated and fall back to the network.

To update to latest (advance the pin, not just re-capture the same commit), bump SELF_VERSION in eng/prepare-decompiler-corpus.sh (and any third-party version), re-emit the real-world baseline, then re-harvest as above. Re-harvest EVIL from its broad pool the same way:

bash eng/prepare-evil-corpus.sh /tmp/evil-pool
dotnet run --project tools/DecompilerHarness -c Release -- \
  --harvest-evil-corpus external/authored-source-corpus/evil/corpus.jsonl \
  --harvest-target 12000 \
  $(cat /tmp/evil-pool/assemblies.txt)

Commit the regenerated corpus.jsonl in the external/authored-source-corpus worktree (it commits to the orphan branch), and land the pin bump on main through a normal reviewed PR. Keep the two in step so the benchmark and the fixed real-world corpus stay aligned.

Benchmark

The benchmark feeds the vendored authored bodies into the same source-correspondence oracle the on-demand census uses and emits a taste card. It validates the corpus and the current decompiler together; run it after a re-harvest or when asked for a decompiler correspondence read:

dotnet run --project tools/DecompilerHarness -c Release -- \
  --benchmark-authored-corpus external/authored-source-corpus/civil/corpus.jsonl \
  $(cat /tmp/corpus-assemblies.txt)

Add --json for structured output. Each target is Correct, one of four valid-but-different taste buckets, Invalid, or a diagnostic bucket. The run exits nonzero on any Invalid, Drift, or Unsupported, and on a dishonest run (unmatched rows or zero targets). A Drift bucket here means the corpus identity no longer resolves in the pinned assembly — treat it like a Drifted verify row.

Guardrails

• Keep the benchmark hermetic: it must stay fully offline over the vendored bodies. Do not add live network resolution to the benchmark path.
• Do not delete the vendor/authored-source-corpus orphan branch or migrate its bodies onto main. The vendored snapshot is the reproducible oracle.
• Build the harness with the solution graph: dotnet build dotnet-inspect.slnx -c Release.
• Verify and benchmark over the pinned assemblies, never a fresh local build of the tool (that reintroduces #1404 circularity).

Report shape

Keep the report short:

Corpus maintenance: <verified | drifted | re-pinned | benchmarked>

- Corpus: CIVIL <verified/drifted/unavailable counts>; EVIL <if checked>.
- Drift cause: <none | upstream churn (re-pin) | outage/--repo (retry) | harness regression (issue)>.
- Action: <none | re-harvest | bump SELF_VERSION + re-harvest | file issue>.
- Benchmark (if run): <correct / taste-bucket / invalid / drift counts>, exit <code>.