feat(scenario): probabilistic conditioning and reverse stress via entropic tilting

[thomaspinder]

Summary

Soft probability targets over existing forecast draws, and the inversion into reverse stress — extending the hard-conditioning stack (ADPRR/Waggoner-Zha) per issue #150:

• ProbabilityTarget / MomentTarget join the scenario vocabulary; ForecastResult.tilt(targets) / ConditionalForecastResult.tilt(targets) return a frozen TiltedForecastResult carrying minimum-KL tilting weights over the parent draws (held by reference), achieved-vs-requested targets, Kish ESS (+ degeneracy warning), and the exact KL divergence. Hard and soft conditions compose by chaining — pins hold pathwise on every draw, so preservation under reweighting is a theorem, not a code path (pinned bit-exactly in tests anyway).
• Solver: exact closed-form two-mass weights for a single probability target (no optimiser); a logsumexp-stabilised convex dual (BFGS + Newton polish with the analytic Hessian = tilted covariance) for the general case, with joint-infeasibility detection (achieved-tolerance + λ-divergence guards) and empty-support errors that count the draws.
• IdentifiedVAR.reverse_stress(variable, threshold, steps, ...): baseline event probability, the tilted distribution, and the most plausible shock cocktail — the tilted-weighted mean of retained structural shock draws, via a new _scenario.structural_forecast_draws that reproduces the scenario engine's RNG stream draw-for-draw (matched-seed nesting pinned bit-exactly under both constant and time-varying volatility, the latter mutation-verified). Plausibility reported as q = ‖cocktail‖² plus q_cal — the ADPRR calibration applied to the tilt's relative entropy, explicitly documented as an extension, not a citation (ADR-0009).
• Weighted median/HDI implemented properly (Hazen-consistent weighted quantiles — exact against np.quantile(method="hazen") under uniform weights; min-width weighted HDI pinned against az.hdi). Fan-chart and cocktail plots included.
• Docs: how-to, ADR-0009, CONTEXT.md vocabulary (with the "importance weights" avoid-term — no proposal distribution is involved), Robertson-Tallman-Whiteman bib entry.

Closes #150

Review

Planned by a Fable-tier planning agent; independently reviewed with the two-mass solution, dual gradient/Hessian, and the RTW Gaussian oracle re-derived from scratch and confirmed numerically. The reviewer endorsed all seven implementation deviations — including the implementer's catch that the plan's own weighted-quantile test was mathematically inconsistent (the mid-point convention is not replication-consistent; resolved via the exact Hazen equivalence). Two P2s (a missing time-varying matched-seed regression pin — verified to hold at diff 0.0 before pinning — and a one-word q_cal doc misattribution) fixed in a follow-up commit.

Tests

75 new tests: exact analytic oracles (two-mass weights/ESS/KL at 1e-13; RTW λ against an independent brentq root-find; dual vs direct SLSQP primal), hard-pin preservation bit-exact, RNG nesting pinned three ways plus the mutation-verified time-varying pin, weighted-quantile/HDI convention pins, and every guard path. Fast suite: 602 passed, 29 deselected. ruff/ty/docs build clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Egjd7ToFeb9TQqFnfRQZxV

[codecov-commenter]

Codecov Report

❌ Patch coverage is 97.57869% with 10 lines in your changes missing coverage. Please review.
✅ Project coverage is 95.3%. Comparing base (866ca8f) to head (f04a7d4).

Files with missing lines	Patch %	Lines
src/impulso/_tilting.py	97.7%	2 Missing and 2 partials ⚠️
src/impulso/_scenario.py	90.4%	1 Missing and 1 partial ⚠️
src/impulso/identified.py	95.2%	1 Missing and 1 partial ⚠️
src/impulso/plotting/_tilted_forecast.py	96.6%	0 Missing and 2 partials ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##            main    #213     +/-   ##
=======================================
+ Coverage   95.0%   95.3%   +0.3%     
=======================================
  Files         45      47      +2     
  Lines       3099    3498    +399     
  Branches     380     415     +35     
=======================================
+ Hits        2945    3336    +391     
- Misses       111     114      +3     
- Partials      43      48      +5     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.