Parameter censoredRunaways: score a runaway replicate as at least the size it burned

[eliotmcintire]

Depends-on: PredictiveEcology/fireSenseUtils#166

A new switch for the spread fit: a simulated fire that reaches the edge of its buffer is scored as "at least as big as it burned" instead of as twice its buffer, which had made the fit prefer no year-to-year variation in fire spread. Off by default; nothing changes until a run turns it on.

Affects: with censoredRunaways = TRUE, fits, their re-scores and the threshold calibration use the censored objective (new cache entries); with the default FALSE every cache key is unchanged.

Technical details

• defineParameter("censoredRunaways", "logical", default = FALSE); fitSpread() passes censoredRunawaysOrNULL(P(sim)$censoredRunaways) to runDEoptim() and estimateSNLLThresholdPostLargeFires() to runSpreadWithoutDEoptim() (both branches hand it to .objfunSpreadFit()), TRUE or NULL; NULL is added to omitArgs through fireSenseUtils::omitNullArgs() in both Cache() calls, as runawayBufferMultiple = NA already is, so a FALSE leaves the keys as they are.
• Needs the censoredRunaways argument of fireSenseUtils (Spread objective: right-censor runaway replicates at the size they burned (censoredRunaways) fireSenseUtils#166); the reqdPkgs floor follows once that merges and the version is bumped.
• Tests: test-objectiveSettings.R (TRUE reaches the fit, the calibration and the objective; FALSE arrives as NULL), test-metadata.R and test-metadataContract.R (name, class, default). 142 module tests pass through the package rendition.
• Evidence (phase4/spreadYearRE in the session logs): on synthetic fire sizes from ELF 14.3's fitted parameters with a known per-year effect of sd 0.25, the objective's minimum over yearSpreadSD was at 0.03-0.11 with the 2 x buffer rule, 0.16-0.18 with censoring (and 0.16-0.19 with penaliseRunaways = FALSE); with the year effect at its true value the optimum of maxAsymptote sits at the truth under censoring and 0.01-0.02 below it under the 2 x rule.

🤖 Generated with Claude Code

https://claude.ai/code/session_014G18Mik3fH4z5M5CyMSUS6

[eliotmcintire]

Closing: Eliot chose penaliseRunaways = FALSE (simulated fires scored at the size they burned, buffer-bounded like observed fires). In the known-sigma tests on ELF 14.3 the censored variant gave the same fits as FALSE within Monte Carlo noise; a larger buffer is the remedy if buffer-filling fires become a problem.