test(code-index): no-call graph rebuild seals zero edges

[ScriptedAlchemy]

Summary

• edge_count 0 with symbol_count 98305 is the sealed relation census of the no-call refresh fixture, not a stalled graph projector.
• The Mac timeout already had that census on the matched generation. A census is only Observed after the graph store is installed, which happens after the full snapshot. The projector had finished.
• Do not change edge emission, and do not raise the 90s RECEIPT_TIMEOUT, in fix(code-index): seat text through retryable graph activation #1562. This follow-up only records the fixture shape.

Citations below are from PR #1562 tip b80dd58 unless the path is unchanged on master (the seat commit does not touch emission). registry.rs and mount.rs line numbers are the seat tip.

Verdict

Expected. Not a stalled projector. Incremental edge emission would not move this receipt.

The status field is graph_statistics.edge_count, which is CodeIndexGenerationStatisticsV1.edge_count: the count of CanonicalRelationEdgeV1 rows sealed with the generation (crates/tracedecay-code-index/src/production/mod.rs:2172-2176, generation_statistics.rs:44-61). It is not a live graph-db relation counter and it does not include CodeFileContainsSymbol projector rows (crates/tracedecay-code-index/src/graph_projection.rs:59).

98305 = 768 files × 128 functions + before_reopen. That is one symbol per function (chunks.rs test now locks the one-function shape). A partial emit would not land on that exact total while still claiming zero call edges and a finished clone/text phase. The number is the sealed generation, not a counter stuck mid-projection.

Why this fixture seals zero relation edges

The batch is:

pub fn refresh_probe_{file:04}_{symbol:03}(input: u32) -> u32 { input + N }

(crates/tracedecay/tests/transport_acceptance_suite/graph_rebuild_status_test.rs:232-237). No calls, no imports, no impls, no modules.

1. The Rust extractor pushes the file node as the parent (crates/tracedecay-code-extraction/src/rust_extractor.rs:166-168) and emits Contains from that parent to each function (280-287).
2. File nodes are not symbol rows (crates/tracedecay-code-index/src/chunks.rs:773-786). canonical_relation_edges abstains those edges as MissingSymbolEndpoint (2497-2502). The existing comment at 4250-4251 states this: the file node is not a symbol row, so its Contains edges always abstain.
3. Parameter and return u32 become unresolved TypeOf / Returns refs (rust_extractor.rs:296-315, 2034-2042). Same-file resolution finds no symbol named u32. Cross-file resolution returns immediately for a bare name with no import and no glob (crates/tracedecay-code-index/src/production/helpers.rs:560-568). Those refs never become edges.
4. collect_edge_evidence only concatenates per-file edges plus cross-file resolutions (helpers.rs:358-371). Both are empty. edges.len() is the census.

A same-file call still seals a Calls edge. The emitter is live. The proof is chunks::tests::no_call_refresh_probe_seals_zero_relation_edges (chunks.rs:4282). Ran:

cargo test -p tracedecay-code-index --lib chunks::tests::no_call_refresh_probe_seals_zero_relation_edges -- --exact

1 passed, 0 failed.

Projector gate, and why it is not this timeout

Graph activation does wait on a full structural backfill before Ready, even when relation edges are empty.

• publish_verified_snapshot builds the whole manifest before any head exists (crates/tracedecay-store-runtime/src/session_registry/code_graph.rs:1233, manifest at 1329-1346).
• build_projection allocates every file entity, every symbol entity, and every CodeFileContainsSymbol relation, then returns (crates/tracedecay-code-index/src/graph_projection/builder.rs:286-384). Symbol hashing is windowed (332-340, workers × 512) but the function does not publish a prefix.
• activate_persistent_graph calls install_graph_serving only after that snapshot returns (crates/tracedecay-code-index-runtime/src/code_index_scheduler/graph_activation.rs:734-774). Ready is that install (serving.rs:3508-3512). Catalog warm is already background and is not the Ready gate (graph_activation.rs:756-758).

For this fixture the rows are about 769 file entities, 98,305 symbol entities, and 98,305 file-contains-symbol relations. The call-edge loop (builder.rs:375-384) is empty. Those containment rows are invisible to edge_count.

The census reader only reports statistics after interactive_graph_store() succeeds (crates/tracedecay-code-index-runtime/src/project_reads.rs:102-119 and 159-174). That store exists only in CodeGraphActivationStateV1::Ready (serving.rs:1993-2011, 2062-2074). An Observed census with symbol_count 98305 means the full snapshot was already installed. edge_count 0 is the sealed answer sitting on a finished projector, not a counter waiting for edges that will never arrive.

progress.phase = ready is the text artifact phase, published when lexical owners install (serving.rs:3206-3211). A clone successor can still start after that. Clone pages 2305/2305 are that successor's source-page observation, not relation emission. Neither input is edge_count.

Ranked proposals

1. Do not land an emission change in fix(code-index): seat text through retryable graph activation #1562. Do not raise RECEIPT_TIMEOUT. Zero relation edges is the fixture. Changing the projector or fabricating containment edges into the census would lie about the source. The proof test belongs in this follow-up, not in the seat-swap PR.

2. Do not stream relation edges to save this receipt. Follow-up only if a different journey measures code_index.seal.collect.emit. There are no relation edges to stream. Streaming the 98k symbol / file-symbol windows would still leave freshness non-current until the last window, because dashboard_terminal_status requires code_graph_serving == Ready (registry.rs:1205-1231 on b80dd58) and Ready is the full snapshot. Estimated speedup on this 90s receipt: none. The dump already showed the finished census. Loosening Ready to the first 512-row window would report a partial graph as terminal; that is not a valid cut.

3. Optional, low, not this timeout: status copy says "graph counts are not authoritative" while warming (crates/tracedecay-mcp/src/handlers/info/status.rs:482-487). That sentence invites reading edge_count 0 as unfinished work. A separate wire field for CodeFileContainsSymbol would stop the misread, but it is a contract change and does not make wait_for_current_generation return. Skip unless an operator surface is actually wrong.

What the receipt is still waiting on

Not this lane, so the next pass does not re-litigate edges.

wait_for_current_generation (graph_rebuild_status_test.rs:153-169) needs freshness status current. That is coverage complete and staleness fresh (status.rs:451-456). The ladder (crates/tracedecay-contracts/src/code_index_freshness.rs:426-466) does not read edge_count. After the seat matches and the graph is Ready, the remaining inputs are reconcile_in_progress, source_verified != Some(false), and hook_hint_count == Some(0), plus source_reference / source_revision.

On b80dd58 the public pass guard is dropped before native graph activation (mount.rs:1159-1173), so a long snapshot does not itself hold rebuild_in_flight. The Mac dump already had seat match, search hits, retrieval serving, age ≈ 89s, clone pages complete, and text phase ready. The next field to read off that same status object is staleness_state, coverage, rebuild_in_flight, and hook_hint_count — not the projector.

Test plan

• cargo test -p tracedecay-code-index --lib chunks::tests::no_call_refresh_probe_seals_zero_relation_edges -- --exact — 1 passed
• No timeout or budget change
• Did not re-run background_refresh_and_reopen_report_only_servable_generations (the receipt under diagnosis; this change does not alter it)

Checklist

• No secrets
• CHANGELOG not updated (test-only proof of an existing census shape)
• No timeout change

  

[changeset-bot]

⚠️ No Changeset found

Latest commit: 11b3cc9

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR