fix(MarketResilience): score non-recoveries, count only the depleted side, drop wall-clock reads

[silahian]

The defect

The Market Resilience tile reported maximum resilience on the worst possible market.

The score has a single write site, reached only when a recovery completes. An event whose liquidity never came back never completed, so nothing was ever written and the tile sat at its cold-start value of 1.0 while the book stayed empty. A book that recovered read 0.31; a book that never came back read 1.00.

What changed

Scoring

• A depth or spread window that elapses without recovery is now scored as a non-recovery: the component reads 0 at its full weight and is not written to the recovery history. Previously the state was cleared and nothing published.
• Only replenishment of the depleted side counts as recovery. Growth on the untouched side is a price move, not resilience. The check that credited it was also degenerate: at activation the untouched side's baseline and floor were set equal, so the denominator collapsed to 1e-9 and any uptick cleared the 90% bar.
• The anchoring trade is now purely the attribution gate. Its expiry no longer clears a live, recovering event — the recovery deadline belongs to the depletion, which is the thing that has to come back. This also removes a depth-sampling stall that occurred during expiry.
• A spread recovery no longer closes the whole event ahead of the depth outcome.
• The first scored event of a session seeds the history and publishes nothing. Previously it compared itself against itself, producing exactly 0.5 on both recovery components regardless of what happened — so a 50 ms and a 700 ms recovery published the same number.

Market-agnostic correctness

• Removed the absolute 0.0001 spread floor, which clamped the magnitude component to 1 on any instrument quoting below it. The component is now omitted when the average spread is not positive. This platform serves crypto, equities, futures and FX; no price-scale constant belongs here.

Determinism

• No wall-clock reads remain in the plugin. The depth timeout ran on machine time while everything else ran on the venue clock, so replay scored differently from live.

Transparency

• Warm-up is disclosed: the calculator exposes its progress and the tile is flagged stale until the depth baseline holds 200 book updates.
• The tooltip described a "Time Recovery" component that has never existed in the code, a three-component model that has been four for a long time, an absolute 0-to-1 reading of a scale that is relative to the instrument's own recent recoveries, and a same-side scoring adjustment that was never built. All corrected.

Bias tile

• Direction now comes from the depleted side that fails to be rebuilt, which follows from the depleted-side rule above.
• The hysteresis latch is evaluated on every scored event, not only one that produced a depth outcome. A spread that widened and came back scores without depleting anything; the latch was checked after the depth guard returned early, so the arrow stayed armed through a recovery that should have cleared it.

Supporting changes

• RollingWindow gains AddWithEviction, returning the item pushed out when the window is at capacity, so running sums can be maintained without re-walking the window.
• StatisticalHelper drops LINQ from its mean and variance paths. Enumerating IEnumerable<T> boxed an enumerator on every update.

Performance

A book update where nothing is happening — the overwhelmingly common case — measured on an Intel Core Ultra 9 285H:

	before	after
depth 10	12,755 ns / 48 B	172 ns / 0 B
depth 25	14,107 ns / 48 B	225 ns / 0 B

The per-callback budget for the shared dispatch thread is 10,000 ns. The cause was the rolling spread statistics running on every update; they now run only while a shock is live.

Deliberately NOT changed

The plugin Description strings. They are hashed into the plugin id that keys saved settings, so editing one orphans every existing user's symbol and provider selection and leaves the tile dead on launch. The corrected wording lives in the tooltip, which is not hashed.

Tests

The existing suite is rewritten where it asserted the old behaviour:

• Tests that fed one event to a fresh calculator and expected a published score now seed a prior recovery first.
• Tests that expected an opposite-side improvement to close an event now let the window close.
• The rewritten tests drive time through the platform's time provider instead of sleeping, so they are deterministic.

No test was deleted. Three were renamed because their names described behaviour that no longer exists, and one was added covering the depth timeout scoring zero at full weight.

Reviewer note

tests/Unit/Studies.MarketResilience.Test executes zero tests on this repository as it stands — dotnet test reports Zero tests ran for every test project here, including ones this PR does not touch. That is pre-existing and independent of this change, but it means CI cannot currently demonstrate these tests passing. The likely cause is that global.json does not pin an SDK version. Worth fixing separately.

[Copilot]

Copilot review overview

🟡 Changes recommended

Empty books can bypass timeout scoring, spread fallback state can be corrupted, and running decimal statistics can overflow on valid large trade sizes.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 1 · 2

Open (3)

• Decimal trade-size squaring can overflow during scoring · New
• Empty book updates bypass timeout handling · New
• Non-positive spreads overwrite the last valid spread · New

What changed in this PR

Updates Market Resilience scoring to handle non-recoveries, depleted-side recovery, deterministic timing, warm-up transparency, and lower callback overhead.

Changes:

• Reworks resilience and bias scoring lifecycle.
• Adds efficient rolling-window statistics.
• Updates tooltips and deterministic tests.

File	Description
PluginSettingsView.xaml	Fixes timeout-label spelling.
StatisticalHelper.cs	Replaces LINQ statistics with loops.
MarketResilienceCalculator.cs	Implements scoring, timing, recovery, and performance changes.
MarketResilienceWithBias.cs	Updates bias direction and hysteresis handling.
MarketResilienceStudy.cs	Adds tooltip, stale-state, aggregation, and lifecycle updates.
MarketResilienceBiasStudy.cs	Updates bias tooltip, filtering, and lifecycle handling.
RollingWindow.cs	Adds eviction reporting for running statistics.
MarketResilienceWithBiasTests.cs	Rewrites bias scenarios with deterministic timing.
MarketResilienceCalculatorTests.cs	Covers new scoring and timeout behavior.
MarketResilienceCalculatorNumericalStabilityTests.cs	Updates numerical-stability expectations.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.