[forhappy]

Summary

Broad questions can miss implementations when they do not name exact symbols, or find a coordinator while omitting its delegated helpers. Add bounded native fielded retrieval, whole-question ranking, source-owned excerpts and an opt-in top-five shortlist shared by CLI and MCP. Eligible questions about one or more operations with only Low-relevance answers can also recover helper inspection leads through exact native call paths. Paths and previews remain explicitly unverified navigation, without transferring caller topics or relevance to a helper.

• Treat ordinary Why/why/WHY as broad-question scaffolding. Previously a capitalized interrogative could become a missing named subject and suppress an otherwise relevant implementation. Preserve quoted/qualified subject guards and exact symbols named Why; refresh disposable responses under ranker29 without rebuilding graphs. This retrieves inspection evidence, not a causal explanation.
• Honor explicit exact source targets. An inadmissible target remains unresolved instead of falling back to an unrelated same-spelled lexical, module or external declaration. This removes fabricated calls and references when a callable parameter shadows an outer function. Missing target IDs still fail ordinary evidence validation. AST cache34 refreshes current disposable graphs; historical realizations remain unchanged.
• Publish immutable source-query/8-callable-comments evidence with up to 64 anchored comments per callable, retaining independent field/file/artifact bounds, ownership, omission counts and historical-profile capabilities. JSON/store retrieval shares the same source evidence and corpus statistics.
• Rank using whole-question coverage, corpus frequencies and complete owned comment passages. Select exact source quotations using action coverage and known positive group frequencies, without borrowing sibling text or manufacturing matches from clipped prefixes. Bounded forward spelling hints retain literal topic guards, explicit mapping disclosure and Low relevance when dependent on hints.
• Recognize ordinary rebuilt/rebuild and written/write spellings while retaining original query groups and explicit identity guards. Keep build/rebuild and write/rewrite distinct; sent/send already uses the existing canonical token profile. These spelling hints preserve action/subject classifiers.
• Preserve discovery's typed exact-name/ID tier in ranked structural answers. Identity labels remain separate from relevance, ambiguity and graph coverage; an exact structural seed can still have Low relevance.
• Continue raw discovery through bounded exact source-witnessed calls. Automatic deeper ordinary context starts only at High-relevance selected roots; explicit inspection retains bounded witnessed paths without promoting relevance. The separate new helper-candidate channel uses source-owned root topics, verified callable endpoints and exact AST call occurrences. Keep explicit-name, negative/temporal, scope, ambiguity and strong-answer guards. Default depth remains2; helper navigation honors the request up to4 hops.
• Retain compact root/ordered-hop IDs, selected call sites and native attribution in callLink, with Low/Approximate helper labels. Only displayed helper paths retain context records after presentation caps, while work counts persist. Show a linked_callable_preview only when a displayed helper has no question-matching quote: direct immutable owner windows, at most16×512bytes/64 comment ranges, source-order ties, and no live-source reads or query-term/ranking changes. Strict consumers reject non-callable preview kinds, inconsistent paths and unsupported profiles.
• Meter additional helper Calls lookup attempts before the backend, including empty/filtered/failed calls: a shared40-probe ceiling beside unchanged indexed recall291. Raw discovery2/work1 and ranked-view7 retain the two ledgers even when no helper survives. These are logical API work counts, not physical I/O or confidence. Bound linked metadata before copying and read store adjacency IDs before hydrating only the retained records; preserve typed limit/corruption errors.
• Correct unconfigured Python module identities: bare roots use relative modules, verified package roots retain their namespace, and configured roots retain precedence/ambiguity. Preserve script/group source identities with dot/percent component escaping and reject unsupported package-entry layouts explicitly.
• Preserve Rust literal/type references, scoped receiver/generic-bound identity, ambiguity and exact occurrences. Explicit deferred receivers cannot become a same-named local method solely because that method is unique.
• Carry supported TypeScript generic arguments through definition-bound imported fields, aliases/substitutions and utility types. Keep export obligations, physical alternatives and original-syntax inventories explicit. Bound proof/load payloads before cloning and charge export walks/fixed-point closure to shared limits. Aggregate overflow retains eligible exact edges through explicit degraded partitioning; incomplete or truncated proofs cannot select a winner. These development controls do not establish compiler-wide package/type/runtime parity or language promotion.

Structural extraction and queries remain native and local, without Graphify, Python, credentials, embeddings, vector databases or grammar downloads at runtime.

Verification

All twelve current native/build/check groups pass on 837 byte-identical Rust/Cargo inputs in the authoritative and build worktrees. Compilation ran only in the build worktree's own external target; this is not a separate authoritative-worktree build:

• Formatting and workspace library/binary Clippy with warnings denied. Both changed test targets (broad_retrieval and code_query_cli) also pass targeted Clippy with warnings denied.
• 1,330 workspace library/binary tests, with two existing ignores across 36 executables.
• 483 query tests across30 executables, including a retained RED/GREEN Why-question regression, case parity, JSON/store equality, reordered graph input and quoted/qualified/exact identity guards. A real-binary CLI regression covers both storage configurations, repeated requests and bounded source-cited text.
• The unchanged exact-binding resolver slice previously passed 661 tests across40 executables at64b4c12c, including its three publication/validation controls and test-target Clippy; these prior results are not claimed as a fresh resolver-suite run for the query-only correction.
• 65 CLI query, 18 MCP query and nine product tests, plus the product-boundary check in both worktrees.
• The complete qualify_code_graph_v1.sh --fixtures-only gate, including lifecycle, scale, semantic/topology, Markdown and nested React checks.

A retained two-file authored fixture verifies the actual CLI transition: the old graph has 24 nodes/40 edges; ordinary update with the new binary removes four false calls and two false references to shadowed outer functions, preserving every node and legitimate edge byte-for-byte. The old binary reuses its cached sources; the new binary reextracts under cache34 without --force. Warm and fresh new builds have identical nodes and edges. Original source, old output, frozen binaries and all native inputs remain intact. This tests denial of unrelated targets; it does not add parameter references or identify their runtime implementations.

All 679 benchmark-harness tests, including 40 focused synthetic conditional-bound controls, pass on unchanged harness source. Scoped product-source reviews found no required correction to the exact-target denial or its cache/validation ownership; they do not certify execution, independent source meaning or comparative quality. Earlier helper/navigation controls remain attributed to their unchanged query source and retained verification. An additional source-only review of the eight Why-question product/test/doc files found no required correction; all frozen hashes match. That review saw no runtime or development-evaluation outcomes and does not certify empirical quality.

The published 64b4c12c CI run completed with 12 successes, including quality, full graph fixtures and all six native platforms, and two dependency failures. It predates this Why-question correction; new-head CI must run after publication. Rustls 0.23.46 clears its four findings. The remaining local policy errors are six ZIP vulnerability records (three advisories on each of ZIP6/7) and unmaintained unsafe-libyaml. Bans, licenses and sources pass; no waiver was added. ZIP7 is transitive through Candle, so upgrading only the direct ZIP dependency is insufficient.

The unchanged broader query-test lint failure at crates/compass-query/tests/explanation_members.rs:289 and historical strict language-promotion failure remain separate from the passing required native baseline. Development producers remain Qualifying. Neither failure is waived. Historical source-review aggregate/incidental disclosures remain recorded; no new real answers or comparison outcomes were supplied to the product-source reviewer. This PR remains draft, with no release or promotion claim.

Quality evidence and remaining work

The prospective developer graph-precision design2/diagnostics2 path adds exact integer hypergeometric tail inversion and simultaneous Bonferroni bounds over captured emitted-record populations. It uses population-size weights, rational difference endpoints, adversarial completion of unknown labels, full withholding after failed builds and bounded arithmetic. Legacy design1/result1 retain their shapes and null certified-interval fields. These calculations are conditional on ideal uniform sampling and truthful labels; the hash sampler does not establish those assumptions or independent review. Design acceptance, certified precision noninferiority, power and the final joint graph/query decision remain unfinished. This is developer tooling, with no Graphify runtime dependency. All679 harness tests, including40 focused synthetic controls, pass for this benchmark component.

Superiority over Graphify remains unproven: zero accepted independent qualification judgments.

A separately authorized evaluator froze six source-first Flask development questions (five supported, one unsupported control) before answers. The Why correction moves the relevant exception handler from an admission filter to rank1 on unchanged graph/source bytes. Required implementation owners in the first five improve 6/10 to 7/10; the other five questions are byte-identical across all15 default/ranked-text/ranked-JSON surfaces. The same frozen fact obligations still yield 0/5 complete behavioral answers: ranked JSON covers3/16 facts fully,6 partly and omits7. The unsupported control makes no invented retry claim but still lacks explicit premise rejection. Two admission failures, a rank10 lifecycle result, quotation omissions, partially omitted long-function windows and a first-line-only Python docstring anchor remain follow-up findings. The full docstring text and correct ownership edge are published; the cause of absent owned-document enrichment is unproven. This small exposed development batch is not held-out or comparative qualification. The owner-binding adapter delimiter correction and one explicitly disclosed evaluator executable-hash read outside the batch directory are retained in the evidence.

Eligible single-operation questions use a bounded lexical operation-head heuristic over the existing owning action classification and full query. Noun-like modifiers removed from navigation hints cannot become independent root-topic credit. Retain up to eight best exact target alternatives per root and display at most two distinct helpers; repeated paths retain one best path, with hint diversity, shortest length and stable occurrence IDs deciding order. More same-operation targets do not alone promote a root. This remains approximate/Low inspection navigation, not English parsing, intent resolution or proof of requested behavior. Current/best root maps can coexist; eight is not a global allocation bound.

The retained own-repository single-operation development replay at 9a79b45b includes validate_with_limits at rank2 with explicit depth4, its exact three-hop source witness and a215byte retained code preview. At default depth2 it reaches the validate wrapper. Adding context preserves the complete ranked response and work counters. These are exposed development observations, not independent answer judgments, general recall, semantic truth or performance comparisons. Native alternative controls separately require two distinct same-operation targets, shortest source paths, stable occurrence selection, backend semantic parity, truthful backend-specific indexed work, and coherent helper-work counters. The earlier detection-vocabulary/partial-filename and context-only experiments remain excluded; optional LSA remains an explicit native opt-in, not a runtime requirement.

Retain the original joint query/graph/cost qualification gates: at least +10 percentage points first-page essential-fact success with paired 95% lower bound above zero; claim-precision and exact/directed noninferiority margin −2 points; independent graph precision/recall, language/relation and occurrence audits, including at least 2000 accepted judgments, precision≥0.995 and Wilson 95% lower≥0.99; and original/lookup/union, output/budget/cost, build/query latency and RSS requirements. A paired graph endpoint, accepted independent source/answer/graph judgments, prospective power and an adequately powered untouched panel remain unfinished. Reverify and freeze the latest Graphify when starting final qualification. The latest development pin is v0.9.92, not a final comparison freeze.

692 observations remain unopened/unassigned and174 previously exposed registered tasks remain excluded. In addition to4 earlier external development answers, this new six-task source-first development batch and its two derived probes are excluded from future held-out claims:36 primary-view captures,36 stage traces and one external graph build. The earlier135 ordinary own-repository captures, one additional warmed JSON-startup diagnostic, and38 context attempts (37 successful/1 historical invalid flag) remain development inputs; the pre-query stale-manifest harness failure is retained separately and executed no query. Authored fixtures are separate from independent judgments. Grep matches, graph size, tests and source approval do not establish semantic truth, independence or comparative superiority. Developer graph helpers retain failures/unreviewed records and emit qualificationEstablished=false.

Compatibility

Source-query profile8 and AST cache34 retain documented rebuild/older-reader requirements; profiles1–7 and historical realizations keep their original meaning. The current broad-query slice uses ranker29, call-link/work1, raw discovery2 when metered, ranked-view7 when metered, and trace2 when metered; unmetered raw1/ranked4/trace1 remain. Experimental unpublished ranked5/6 are unsupported. Strict consumers must upgrade for the new raw/ranked majors and validate both work ledgers. The generic Agent View shape remains major1 with the actual raw identity and an informational work caveat; canonical counters remain in raw stats.

Ranker29, quotation epoch6-linked-preview and call-link epoch7-classified-operation-heads refresh disposable responses. Existing spelling/identity behavior remains. Query-only changes require no source rebuild by themselves; missing graph/source evidence still requires compass update <root>. Compatibility/reference docs, migration guidance, help/MCP descriptions and changelog specify the new profile/preview bounds. Operational stats remain outside semantic result digests; a raw major still distinguishes metered responses. No immutable historical graph is rewritten.

Python 2, TypeScript 5/JavaScript 4, Rust 4 and Go 2 development producers remain unpromoted. Evidence/2 includes the documented typed argument/load/inventory capabilities; strict older readers must upgrade. Fact-digest profile 4 refreshes current disposable state without rewriting history. Graph-v1 shape and provenance remain intact. Compatibility/reference docs, migration guidance and changelog describe the relevant changes. Local evidence, generated graphs, competitor artifacts, machine configuration and private source content are excluded.