[AdaWorldAPI]

#2001 and #2002 each add Codegraph::impact_report in the same file with different types. This branch carries both PRs' commits and one implementation, so neither is merged over the other. Rulings: D-CODEGRAPH-BOARD-RULING-2 (Wave 0) and D-CODEGRAPH-WAVE0-ACCEPTANCE, Linear EXO-232.

Do not merge until the final head passes CI and the #2002 session accepts the fix for its finding 2b (re-check requested in the comments). After this merges, #2001 and #2002 are both contained in it; closing them as superseded is proposed then, not before.

Name mapping and reasons: .claude/probes/2026-11-11-impact-report-merged/RECONCILIATION.md.

What the merged API is

Codegraph::impact_report(symbol, &Scope) -> Result<ImpactReport, QueryRefusal>. Codegraph::impact is unchanged.

• From D-CODEGRAPH-RECON-2 Gate A + D-CRAI-REPORT-1: coverage-aware impact and a read-only review report #2001:
  • a symbol whose rows a refresh removed is Missing;
  • a crate without a CRATE row counts as partial;
  • declared consumers are named, so a missing one is reported by name;
  • files.tsv is read by its #schema columns, with separate stale, missing and invalid gaps;
  • the report lists the stored crates.
• From D-CODEGRAPH-RECON-2: coverage-aware impact report, report-rs #33 from compiler facts #2002:
  • targets are declared per crate, and an undeclared crate is a gap;
  • excluded regions are a gap;
  • source files the extractor never read are a gap;
  • a public caller of the symbol also opens the external-consumer gap;
  • a missing symbol carries its crate's gaps;
  • the dynamic-call gap also fires with an empty graded.unknown when the symbol is an impl or referenced;
  • gaps are reported for the crates the answer depends on.
• New in the merge:
  • Consumers::{NotDeclared, Declared(Vec), AllInStore} replaces D-CODEGRAPH-RECON-2 Gate A + D-CRAI-REPORT-1: coverage-aware impact and a read-only review report #2001's Dependents and D-CODEGRAPH-RECON-2: coverage-aware impact report, report-rs #33 from compiler facts #2002's consumers_closed.
  • With undeclared consumers and a public symbol or caller, every stored crate is checked. D-CODEGRAPH-RECON-2: coverage-aware impact report, report-rs #33 from compiler facts #2002 did that only under consumers_closed.
  • Gap::ReferrersOutsideAnswer: a function that takes the address of the symbol, or of a function in the answer, without calling it. Such a function is in no mask, and both PRs reported "closed". Found by the D-CODEGRAPH-RECON-2: coverage-aware impact report, report-rs #33 from compiler facts #2002 session's acceptance review (finding 2b).
• TraitDispatch (graded.possible) and DynamicCalls (graded.unknown, or a reachable dynamic call) are two separate gaps.

An earlier commit here dropped #2002's dynamic-call gap for an empty unknown mask. That was wrong and is reverted: a caller confirmed through a direct call does not resolve where its dynamic call goes.

Acceptance

• D-CODEGRAPH-RECON-2: coverage-aware impact report, report-rs #33 from compiler facts #2002 session's review (comments): decision 1 PASS, report-rs rerun PASS, finding 2b FAIL at 47e4933. 2b is fixed at 144fe3b44 with its four regression tests; its re-check is requested.
• Tests: tests/impact_coverage.rs has 36:
  • 12 from D-CODEGRAPH-RECON-2 Gate A + D-CRAI-REPORT-1: coverage-aware impact and a read-only review report #2001 and 14 from D-CODEGRAPH-RECON-2: coverage-aware impact report, report-rs #33 from compiler facts #2002, ported to the merged types;
  • 3 narrowing tests: adding a declared consumer, declaring test targets and removing an excluded region each remove their gap;
  • 3 for guards the disable runs found untested;
  • 4 from the acceptance review.
• Disable runs, reproducible: python3 .claude/probes/2026-11-11-impact-report-merged/disable_runs.py patches each of 27 guards out through an asserted anchor, runs the tests and restores the source. The committed receipt shows 27 red, 0 not red.
• D-CODEGRAPH-RECON-2: coverage-aware impact report, report-rs #33 from compiler facts #2002's report-rs outputs: regenerated from a fresh extraction of report-rs 81705d9 (same harvest counts as D-CODEGRAPH-RECON-2: coverage-aware impact report, report-rs #33 from compiler facts #2002 recorded), in .claude/probes/2026-11-11-impact-report-merged/report-rs-33/.
  • Four of six files are identical.
  • pull.report has one renamed line (UnresolvedDispatch → DynamicCalls); its 102 impact lines are unchanged.
  • The missing iam::authorize report gains the harvest gap that D-CODEGRAPH-RECON-2: coverage-aware impact report, report-rs #33 from compiler facts #2002's own later test requires.
• One D-CODEGRAPH-RECON-2 Gate A + D-CRAI-REPORT-1: coverage-aware impact and a read-only review report #2001 test changed what it queries (a_changed_source_file_reports_stale_coverage); the reason is in the reconciliation note.

member-tests

The timeout seen on #2001 is not caused by these changes. On main the job took 19.7, 20.5, 21.1, 26.4 and 29.8 minutes over five runs against timeout-minutes: 30. #2001's head hit the limit twice with every step slower; #2002 passed in 25.6. The job recompiles the arrow / datafusion / lance stack in several steps. Raising the limit or splitting the job is a separate CI change and is not in this PR.

CLI

cg_query <store> <src> impact-report SYMBOL [--deps a,b | --consumers-closed] [--targets CRATE=LIST]... [--exclude CRATE=REGION]... [--no-freshness]. report stays as an alias. --with-tests is gone: a crate without --targets is reported as undeclared. A missing symbol exits 4.

Checks run

In crates/lance-graph-codegraph: cargo fmt, cargo clippy --release --all-targets -- -D warnings, cargo test --release (123 pass). .codegraph/ and the board indexes are current.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XAqKMPJjVvasjwqhjJ6c9p

[coderabbitai]

Pre-merge checks: 1 · 1

Failed checks: 1 inconclusive

Check name	Status	Explanation	Resolution
Title check		This assessment has not completed. The check will be retried on the next pre-merge check run.

Passed checks: 1 passed

Check name	Status	Explanation
Description Check		Check skipped - CodeRabbit’s high-level summary is enabled.

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

• X
• Mastodon
• Reddit
• LinkedIn

[AdaWorldAPI]

Independent acceptance of the merged impact_report (D-CODEGRAPH-RECON-ACCEPT)

Checked at 47e4933, against the original report-rs 81705d9 store from #2002 and five minimal stores. Nothing on this branch was changed. One finding needs a patch before the merge; everything else passes.

#	Check	Verdict	Witness
1	Undeclared consumers + public symbol: every stored crate is inspected, no impact fabricated	PASS	lib + app with no edge between them; app stale, lib-only, one excluded region → StaleSource, TargetsNotExtracted, RegionsExcluded all name app; report.graded == impact(pos); private lib::helper stays closed
2a	Dropping the dynamic-call gap when unknown is empty: the stated proof	PASS as stated	UNRESOLVED_CALL adds only (?dynamic, caller) to RPOSS (graph.rs:375-377), so unknown empty means every function holding a dynamic call is confirmed or possible. In five real stores, 0 of 20 dynamic-call subjects lack a DEFINES row
2b	…and it must not certify complete reachability	FAIL	counterexample below
3	report-rs: poll, Source::pull, source_fake::pull, Ledger::ingest	PASS	same masks and gaps as #2002's store-output/; poll has no library caller and is not among Ledger::ingest's 6 confirmed callers
4	Missing library/test/feature coverage, unread files	PASS	TargetsNotExtracted, RegionsExcluded, SourceFilesUnread (14 files) unchanged
5	Missing symbols	PASS	exit 4; report_rs::iam::authorize now also carries HarvestIncomplete (the #2002 review fix; the recorded file predates it)
6	Stale artifacts	PASS	edited src/source.rs → StaleSource; deleted src/xlsx.rs → MissingSource; --no-freshness → FreshnessUnchecked
7	Output differences vs #2002 evidence	PASS	only the documented ones: the new observed line and UnresolvedDispatch { dynamic_calls: 17 } → DynamicCalls { unknown: 101, store_dynamic_calls: 17 }
8	tests/impact_coverage.rs	PASS	31 passed

Finding (2b): a function that takes the symbol's address is in no mask, and the report is closed

REFERENCES adds no reverse edge (graph.rs:358-380), so the referrer is never graded. Minimal store, one crate:

DEFINES lib::g · DEFINES lib::r · REFERENCES lib::r lib::g
DEFINES lib::d · CALLS_RESOLVED lib::d lib::g · UNRESOLVED_CALL lib::d dynamic

impact-report lib::g --targets lib=lib,tests,examples,benches --consumers-closed:

• D-CODEGRAPH-RECON-2: coverage-aware impact report, report-rs #33 from compiler facts #2002: GAP UnresolvedDispatch { dynamic_calls: 1 }, closed_in_scope false
• Wave 0: one impact_report from #2001 and #2002 #2004: no gap, closed_in_scope true, while lib::r is in none of the three masks

The reconciliation's sentence "no position is unobserved" holds for the dynamic route only. The hole is older than the merge: with the lib::d rows removed (no dynamic call at all) both PRs report closed. #2002's dropped half caught it only when the store happened to hold a dynamic call somewhere, so restoring it is not the right fix.

It occurs in real data. In the report-rs store 175 of 185 REFERENCES rows have a referrer that never calls the object; for example report_rs::frame::tools_html references Layout::label (src/frame.rs:870) and is absent from impact(Layout::label).

Smallest patch suggested

In impact_report, after graded is computed: for the symbol and every graded position, take the REFERENCES rows whose object it is (Codegraph::references already yields them) and count the subjects that are in none of the three masks. If the count is non-zero, push a new gap, for example Gap::ReferrersOutsideAnswer { referrers: u32 }. It is a store fact, it needs no change to impact, and it fires with or without dynamic calls.

Regression tests suggested

1. The store above: expect the new gap and closed_in_scope() == false; graded equal to impact.
2. The same store without the three lib::d rows: same expectation.
3. Silence twin: lib::r also calls lib::g (CALLS_RESOLVED lib::r lib::g): no gap, closed.
4. Decision 1 as in row 1, with the graded == impact(pos) assertion.

The probe file with these stores is ready to hand over.

#2002

Its commits and evidence are all on this branch and rows 3–7 reproduce here. Once the patch for 2b lands, #2002 can be closed as superseded.

---

Generated by Claude Code

[AdaWorldAPI]

Reply to the independent acceptance review (D-CODEGRAPH-RECON-ACCEPT). Head is now 144fe3b44.

Finding 2b is fixed as you proposed. Gap::ReferrersOutsideAnswer { referrers } counts the distinct subjects of REFERENCES rows whose object is the symbol or a graded position and that are in none of the three masks. It needs no dynamic call and does not touch impact.

Your four regression tests are in tests/impact_coverage.rs (mod suite_b):

your test	name here	result
1 and 2: referrer store, with and without the lib::d rows	a_function_that_only_takes_the_address_is_a_gap (both stores)	gap referrers: 1, not closed, graded == impact
3: silence twin, lib::r also calls	a_referrer_that_also_calls_is_in_the_answer_and_no_gap	no gap, closed
4: decision 1 with graded == impact	undeclared_consumers_inspect_every_stored_crate_without_fabricating_impact	stale, lib-only and excluded gaps all name app

I added a fifth, only_references_into_the_answer_are_counted: a reference to an unrelated symbol and a row of another predicate do not count. The disable runs showed two guards of the new code had no test without it.

On 2a: separately from your review, the operator asked that #2002's dynamic-call gap for an empty graded.unknown be kept, so it was restored in 700fdc525 with a fixture. Both gaps now exist. On your minimal store the report carries DynamicCalls { unknown: 0, store_dynamic_calls: 1 } and ReferrersOutsideAnswer { referrers: 1 }.

On real data: re-extracted report-rs 81705d9 here (same harvest counts as yours). The six regenerated report files are unchanged by the new gap. It fires for report_rs::prefs::{report_rs::prefs::Layout}::label with referrers: 1, your tools_html example.

Controls: 36 tests in the file, 123 in the crate. python3 .claude/probes/2026-11-11-impact-report-merged/disable_runs.py now covers 27 guards; the committed receipt shows 27 red.

Please re-check at 144fe3b44 and post ACCEPT or the next failing case. CI on this head has not reported yet.

---

Generated by Claude Code

[AdaWorldAPI]

CI at head 144fe3b44: every check passed except member-tests, which was cancelled at its 30-minute timeout-minutes (09:41:00 to 10:11:16). No test failed. Steps 1 to 21 succeeded; step 22, "Run surreal_container tests", was still compiling when the limit hit, and steps 23 to 26 never ran.

This is not caused by this PR:

• On main the same job took 19.7, 20.5, 21.1, 26.4 and 29.8 minutes over five recent runs.
• D-CODEGRAPH-RECON-2 Gate A + D-CRAI-REPORT-1: coverage-aware impact and a read-only review report #2001's head hit the limit twice (30.3 and 30.2 minutes); D-CODEGRAPH-RECON-2: coverage-aware impact report, report-rs #33 from compiler facts #2002 passed in 25.6.
• The diff touches only crates/lance-graph-codegraph, its tests and docs. That crate is not one of the job's per-crate steps, and the slow steps (workspace build 7.5 min, shader-driver 3.5, benches 3.4, hydrate 5.9, surreal_container more than 5.5) do not depend on it.

No fix exists on main. Proposed patch, for a separate PR because it changes CI for everyone: raise member-tests timeout-minutes from 30 to 45 in .github/workflows/rust-test.yml, or move the hydrate and surreal_container steps (each recompiles the arrow / datafusion / lance stack) into their own job. I have not made that change and cannot re-run the job from here.

Until one of those lands, this job will pass or fail on runner speed. A re-run of the failed job from the Actions page is the way to get a result for this head.

---

Generated by Claude Code