fix: leave receiver-blind calls unresolved #398
No reviewers
Labels
No labels
code-review
correctness
dos
performance
security
severity/high
severity/low
severity/medium
tech-debt
Kind/Breaking
Kind/Bug
Kind/Documentation
Kind/Enhancement
Kind/Feature
Kind/Security
Kind/Testing
Priority
Critical
Priority
High
Priority
Low
Priority
Medium
Reviewed
Confirmed
Reviewed
Duplicate
Reviewed
Invalid
Reviewed
Won't Fix
Status
Abandoned
Status
Blocked
Status
Need More Info
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set
Reference
h-dv/code-index!398
Loading…
Reference in a new issue
No description provided.
Delete branch "codex/csharp-extension-safety"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
An instance call such as
widget.Describe()could resolve to an incompatibleDescribe(this string)extension when receiver origin was unavailable. Removing that extension exposed another wrong-target path through unique method-name proximity or namespace imports. The runtime now leaves these calls extracted and explicitly unresolved.Remove the name-only extension tier and gate the unique-candidate tiers using the existing C#/PHP receiver-locality policy. Lexical/self calls and origin-backed lookup retain their paths. Historical provenance code 70 remains decodable. Schema 86 rebuilds disposable caches so old false edges cannot survive an upgrade. Extension declaration metadata, capability dependencies, guest ABI, SDK inputs, and all eleven provider package identities remain unchanged. Runtime 0.41.3 uses the existing C# 0.1.5 provider; no provider release cycle is needed.
There is a real recall cost. Exact call-site comparison of signed v0.41.2 and the candidate on identical published package bytes withdraws 232 Dapper and 903 Guzzle method bindings, with no target gains, target swaps, or extraction drift. The other five tier-one languages retain their binding baselines. Some withdrawals are correct static/instance/extension calls without sufficient origin proof, including Dapper's 18
PocoData.ForTypesites. Wrong targets removed include Mighty model calls bound to benchmark methods and Guzzle response-interface calls bound to a test-only request implementation. Safe recall restoration requires provider-owned lookup evidence and instance precedence (#313).Dapper's hand-verified benchmark still returns 53/62 answers, with the same nine known misses. Twenty answers now arrive as explicitly unproven name fallback: SanitizeParameterValue 3, LookupDbType 1, SetupCommand 10, CastResult 3, and the same three Sanitize sites scored again by rename_edit_site. The exact fallback-channel pin moves 5 to 25; oracle truth, recall floors, cost records, attribution, and tolerances stay unchanged. The injection fixture uses an explicit static import for its two positive calls and retains the distant qualified call as a negative control.
CI also exposed a Windows permission-refusal timing failure: one call took 2.061 s against its 2 s bound. The permission probe repeatedly paid SQLite's Windows access-denied retry loop. Windows now uses the existing direct file probe; Unix keeps SQLite-first probing to preserve POSIX locks. No caching, new dependency, permission-policy change, or relaxed bound. The cross-process lock-preservation test now runs on Windows too and accepts only SQLite BUSY/LOCKED as proof.
Removing the extension tier eliminates 21 SQL statements even for non-C# projects. Before/after statement traces attribute about 34,000 Express and 32,000 Sinatra full-scan steps to those retired statements; both also lose two sorts and keep the same automatic-index work. The guarded seven-repository cost writer lowers the cost record with the existing tolerance policy and no degraded-stage allowance. The record also discloses prior in-band drift separately from this patch's exact trace comparison.
Validation:
widget.d7d8c9d0passed the Linux ordinary/timing/artifact/doc job and the complete native Windows lane (4,761 ordinary tests plus timing checks, including the permission regression and new lock-preservation case). Final changes after that are the corpus fixture and reviewed cost record only.C# still has static/instance/extension recall gaps, and the plugin ABI/package format remains experimental. This PR does not claim complete C# production readiness.
Fixes #397.