W5 — Tests, Gates, Documentation

Goal: make the defect classes in W1–W4 impossible to reintroduce unnoticed. Owner decisions, 2026-10-02: fix all 15 build warnings, then enable TreatWarningsAsErrors. Status: 3 / 7 done (W5-01, W5-04, W5-05)

---

W5-05 · "Async fake" test provider · P0 · ✅ done#

Problem. Every debate test runs on FakeLLMProvider, which completes synchronously. Any code that is correct only while nothing suspends therefore passes CI and fails in production. This is not hypothetical: it is exactly how W1-01 shipped — 11 green streaming tests, and a feature that returns nothing against a real provider.

The existing FakeLLMProvider already accepts chatDelayMs; the problem is that nothing makes asynchronous completion the default expectation.

Fix (done). FakeLLMProvider in both test projects now suspends by default: an await Task.Yield() before completing, with an explicit suspends: false opt-out and a Suspends property a test can assert on. The class doc carries the incident so the reason survives.

The rule is written where a contributor will actually read it — this repository has no AGENTS.md, so it went into CONTRIBUTING.md under a new "Testing rules" heading, next to the build-verification instructions. The same edit corrected two stale claims in that file: it claimed .NET 8 / C# 12 (the projects are net10.0 with LangVersion preview) and a uniform four-space indentation (Core uses three).

Acceptance.

• The whole suite passes with an always-suspending provider — nothing in it depended on synchronous completion, which is the result worth having
• The rule is in CONTRIBUTING.md
• Both fakes document the incident and the opt-out

---

W5-01 · Zero warnings, then TreatWarningsAsErrors · P1 · ✅ done#

Problem. 15 warnings on a clean build. None are cosmetic:

Warning	Location	Nature
CS9113	Voting/IVotingStrategy.cs:143	dead public parameter (W3-01)
CS8604, CS8602, CS8604	Memory/IAgentMemory.cs:215,292,294	nullability in a public API (W3-03)
CS0169	Council/CouncilBuilder.cs:45	dead field (W3-02)
CS8600, CS8604	Redis/RedisDebateCache.cs:58	null JSON deserialised into a non-nullable
CS8601	Server/Mcp/DeliberaMcpTools.cs:91	null assigned to a non-nullable string (W4-09)
CS9113	tests/…/AdaptiveStrategyTests.cs:311	unread parameter
ASPDEPR002 × 4	Server endpoints	deprecated OpenAPI API (W4-01)
NU1902 × 2	Microsoft.Build.Tasks.Git	vulnerable dependency (W0-02)

Fix (done). All 15 are gone; the build is clean even with -warnaserror:

• WeightedVotingStrategy(defaultWeight) removed (W3-01) — the parameter was never read
• MemoryEntry metadata normalised at both recall sites (W3-03) — line 292 was a latent NullReferenceException, not just a warning: a vector-store hit without tags would have taken pgvector-backed recall down
• RedisDebateCache uses value.ToString() instead of the possibly-null (string) cast
• DeliberaMcpTools passes strategy ?? "Standard" (W4-09) and serialises its error payloads
• CouncilBuilder._persistedOptionsSnapshot deleted (W3-02); unread test parameter removed
• ASPDEPR002: .WithOpenApi() calls deleted (W4-01)
• NU1902: Microsoft.SourceLink.GitHub 10.0.301 → 10.0.303 (W0-02)

TreatWarningsAsErrors is applied in CI via dotnet build -warnaserror, leaving local iteration warning-tolerant.

Acceptance.

• dotnet build -c Release → 0 Warning(s)
• CI fails on any new warning

---

W5-02 · Tests for code that has none · P1 · ⬜ todo#

Problem. Measured gaps (0 references in the entire test project):

• Chunking/AutoChunker.cs (498 lines) and AutoChunkingOrchestrator.cs (255) — the progressive-disclosure feature has no tests at all
• Compression/CompressionCache.cs (240) — no tests, although release notes advertise IDisposable
• DependencyInjection/ServiceCollectionExtensions.cs (340) — no tests, which is why W1-02 (IMemoryCache never registered) was never caught
• CouncilBuilder.ApplyOptions — only indirectly covered
• Benchmarking/CouncilBenchmark.cs — exists, never run

Fix. Prioritised by blast radius: DI extensions first (they are the entry point for every consumer), then AutoChunker, then CompressionCache.

Acceptance.

• UseInMemoryCache / UseFileCache / AddDelibera have resolve-level tests
• AutoChunker has boundary tests (empty input, one-chunk, forced multi-chunk)

---

W5-03 · Concurrency tests for the claims made in release notes · P1 · ⬜ todo#

Problem. v10.3.0's release notes advertise ConcurrentDictionary, volatile int status, timer-based eviction and IDisposable caches. There is no test that exercises any of them under concurrency; LocalDebateOrchestrator has five tests covering only "unknown id" and "duplicate id".

Fix. Tests for: timer eviction of completed entries, Dispose while a debate is active, concurrent GetStatusAsync visibility of a status change, and the W1-05 round collection under concurrent read/write.

Acceptance.

• Each advertised concurrency guarantee has a test that fails without the guarantee
• No flaky test: run the concurrency suite 20× in CI

---

W5-04 · CI on every PR · P1 · ✅ done#

Problem (worse than expected). The workflow existed and ran on pull requests, but:

1. it built src/Delibera.slnx, which lists only the four src/ projects — the test projects were never compiled or executed in CI at all;
2. it never ran dotnet test.

That is the mechanical reason the NU1605 drift of W0-01 could ship: nothing in CI ever built tests/.

Fix. The workflow now restores and builds the root Delibera.slnx (the one that includes tests) with -warnaserror, then runs dotnet test and uploads a TRX report. Publishing stays tag-gated.

Still open for the owner: the repository contains two solution files — the root Delibera.slnx (src + tests) and src/Delibera.slnx (src only). CI now points at the root one, but the duplicate is a trap for anyone who types the obvious path. Deleting it is a repository-layout decision, so it is left in place and flagged here.

Acceptance.

• A deliberately broken build fails the workflow
• Release packaging is tag-gated

---

W5-06 · CHANGELOG · P1 · ✅ done#

Problem. The changelog is the only place a consumer learns about the breaking changes in this cycle.

Acceptance.

• Every breaking change since 10.3.0 is listed with its replacement — ## [10.3.0] carries the P-01 table, ## [10.4.0] carries W3-07 with before/after code, and [10.5.1] states plainly that it introduces none
• The 10.3.0 entry's claim of "402 unit tests pass" is reconciled with reality — the claim was never actually written into the entry, so there is nothing outstanding to correct. Worth recording rather than leaving implied: the same class of drift did reach nuget.org through PackageReleaseNotes (the 514 figure), and that one is fixed and called out in [10.5.1].

---

W5-07 · Documentation sync · P2 · 🟡 partial#

Problem. docs/ is extensive (QuickStart, Server, caching, WhatsNew × 3, NET10 upgrade × 2, ROADMAP) and drifts: several documents describe behaviour this plan is changing, and the deleted docs/NET10-Upgrade.pdf files are still tracked in git.

Acceptance.

• Git index matches the intended file set — the two docs/NET10-Upgrade.pdf files are no longer tracked, and no document outside this one still references them
• Behaviour-changing documents updated for 10.5.1 — QuickStart (EN/RU), Server.md, caching.md, distributed-debates.md, src/README.md, src/Delibera.Core/README.md, README.md, README-RU.md, ChatClientLLMProvider.md, CONTRIBUTING.md, plus new docs/WhatsNew-v10.5.1.md
• Every documented code sample compiles — not verified. The samples in the files changed for 10.5.1 were written against the source and every With… method they name was checked to exist, but no documentation-wide sample compilation was run. Treat this row as open.

---

W5-08 · No framework regression in the debate loop · P1 · ✅ done#

Problem. Nothing compared framework cost between releases, so a change that made every debate ~2× slower to orchestrate would ship green. That is exactly what happened in 10.5.1: cost accounting ran unconditionally, concatenating the prompts and running the token counter over prompt and response on every member call — a cost quadratic in the debate's own length, for a result that was then discarded.

Measurement. .bench/PerfCheck: a fake provider returning instantly, three members, three rounds, long deterministic responses, and the baseline checkout supplied as a git worktree through -p:DeliberaCorePath=. Wall time against live models is useless here — the proxy on this machine dominates latency by orders of magnitude.

Checkout	Median	p95	Allocations
v10.5.0 (1297e38)	0.634 ms	1.22 ms	3 735 640 B
v10.5.1, unguarded (978f4cb)	1.404 ms (+121%)	1.96 ms	5 034 080 B (+35%)
v10.5.1, guarded	0.679 ms	1.41 ms	3 736 000 B

Fix (done). Accounting runs only when DebateExecutionOptions.CostTrackingEnabled. A second pass removed an async ValueTask from the per-member admission check — its state machine was allocated on every call even though the method returned immediately.

Acceptance.

• Allocations within noise of the previous release (+0.01%)
• Wall time within the noise band — repeated 80-iteration runs put the two within each other's spread (0.52–0.60 ms vs 0.56–0.66 ms), so the residual difference is not treated as signal
• Cost gating still works where configured — CostGateIntegrationTests covers both the enforcing and the permissive case

---

W5-09 · Indexing produces the same corpus regardless of run count · P1 · ✅ done#

Problem. W2-15: every index run appended, because point ids were random. The 10.5.0 measurement found 72 points for 24 unique chunks.

Measurement. Live Qdrant, same corpus and queries, varying only how many index runs had happened.

	v10.5.0	v10.5.1
3 index runs of one document	360 points	120 points
Search median afterwards	0.968 ms	0.873 ms

Fix (done) — deterministic point ids; see W2-15.

Acceptance.

• Point count independent of run count
• Latency improves as a consequence
• Ranking quality under duplicates improves — not measured. The latency win is modest because a few hundred vectors stay index-resident; the real effect is on top-k ranking, and demonstrating it needs a recall/precision comparison this harness does not perform.

Why this gate exists at all. W5-05 closed the fake-vs-real gap for behaviour. This is its performance twin: a fake provider returning instantly is exactly what makes framework overhead measurable, and a feature that made the pipeline twice as slow passed every behavioural test.