Remove fit_power_curve.R: PerturbPlan's closed form is the analytical estimate

docs/pysceptre-backend.md

@@ -45,7 +45,8 @@ them needs R.
 | `src/run_power_simulation.R` + `lib/simulate.R` + `lib/pert_input.R` | **ported to Python**; the NB draw, guide-to-guide variability, centering and seeding are WattEG's method and stay in WattEG (§6) |
 | `src/consolidate_replicates.R` | ported (pyarrow) |
 | `src/compute_power.R` + `lib/stats.R` | ported (Wilson interval) |
-| `src/summarize_power.R`, `src/fit_power_curve.R` | ported |
+| `src/summarize_power.R` | ported |
+| `src/fit_power_curve.R` | **deleted** (2026-09-28): the analytical estimate is PerturbPlan's closed form, `pysceptre.analytical_power` |
 | `patches/`, `lib/apply_patch.R`, `src/check_sceptre_api.R`, `src/install_sceptre.R`, `src/install_ondisc.R`, `src/audit_dependencies.R` | **deleted.** All six exist only because the pipeline reaches into unexported sceptre S4 slots and patches its CRT path |
 | `lib/sceptre_io.R` | **deleted from the pipeline**; its odm-materialisation logic moves to the one-off export (§4) |
 | `src/make_test_data.R` | ported, or replaced by a synthetic fixture generated in Python |
@@ -655,8 +656,8 @@ one stage with an absolute bar rather than a relative one.
    `compute_power` and `summarize_power`, checked against R **on identical input**, which makes
    them exact comparisons rather than statistical ones: every column agrees to machine epsilon,
    in R's column order. The comparison caught `max_effect_size_tested` missing entirely and the
-   per-gene columns sitting in the wrong place. `fit_power_curve` is **not** ported — it serves
-   the paper's reduced-design study rather than the pipeline, and nothing in the DAG calls it.
+   per-gene columns sitting in the wrong place. `fit_power_curve` was not ported, and was
+   deleted on 2026-09-28; nothing in the DAG called it.
 6. **Nextflow rewiring** — done: `FIT_NULL_MODELS` and `MERGE_NULL_MODELS` are gone, the six
    remaining processes call the `watteg-*` entry points, the samplesheet column is `dataset`
    (a `.h5mu`) rather than `sceptre_object`, and `pixi.toml` holds no R. **Still open: the new
@@ -825,8 +826,10 @@ figure always clears the 4 GB floor), while the 1,000 Python cis tasks peaked at
 The largest, simulating from `exp(X . beta)`, has been taken; the other two are judgement calls
 about which cells and which estimator belong, and both change every published number.
 
-**One thing deliberately not ported.** `fit_power_curve.R` serves the paper's reduced-design study
-rather than the pipeline, and nothing in the DAG calls it. It stays in R.
+**One thing not ported, and since deleted.** `fit_power_curve.R` fitted a per-pair probit power
+curve for the paper's reduced-design study; nothing in the DAG called it. It was removed on
+2026-09-28: the analytical estimate is PerturbPlan's closed form (`pysceptre.analytical_power`),
+and the simulation stays the reference.
 
 **The synthetic fixture.** `src/make_test_data.R` produces a sceptre object; the Python path needs
 a `.h5mu`, so `assets/samplesheet_synthetic.csv` points at a file nothing generates yet. The stub

docs/status.md

@@ -941,20 +941,12 @@ published since. This one happened to come back clean; the next one may not. Pin
 their locked versions would make future re-solves honest, and is worth doing before the next
 dependency change rather than during one.
 
-### Step 11 — fit the power curve and run three effect sizes instead of six
-
-`src/fit_power_curve.R` fits each pair's power curve so a sweep needs three effect sizes rather
-than six, and makes the minimum detectable effect size continuous rather than snapped to whichever
-effect sizes were run. Usage is in [Usage]({{ site.baseurl }}{% link usage.md %}); the model is
-derived in [Methods]({{ site.baseurl }}{% link methods.md %}).
-
-**The validation — whether three points really do reproduce six, the covariate model that predicts
-the curve for pairs never simulated, and the calibrated prediction intervals — moved to
-[broadinstitute/WattEG-paper](https://github.com/broadinstitute/WattEG-paper) on 2026-09-04**, with
-the rest of the paper analyses. Short version of what it established, so a pipeline user knows what
-the feature is worth: the best three-point design reproduced the six-point per-pair MDES for 92.0 %
-of pairs on `day0`, against a reference whose own bootstrapped self-agreement was 91.4 %. Grid
-placement is dataset-specific and a one-effect-size pilot is enough to choose it.
+### Step 11 — fit the power curve (removed)
+
+`src/fit_power_curve.R` fitted each pair's power curve so a sweep could run three effect sizes
+rather than six. It was removed on 2026-09-28, with the rest of the project's own power model:
+the analytical estimate is PerturbPlan's closed form (`pysceptre.analytical_power`), and the
+simulation stays the reference. The fit and its validation are in git history.
 
 ## Running the comparisons on the cluster
 

src/summarize_power.R

@@ -126,8 +126,9 @@ for (column in shared) {
 
 # Per-gene values from sim_input, joined here rather than carried through the simulation.
 #
-# WHY HERE. The theory behind the power curve is SE^2 ~ (1/n_pert_cells) * (1/mu + 1/theta), so any
-# covariate model of power needs the gene's dispersion as well as its expression. Only
+# WHY HERE. The test statistic's variance goes as SE^2 ~ (1/n_pert_cells) * (1/mu + 1/theta), so an
+# analytical power estimate (PerturbPlan's closed form) needs the gene's dispersion as well as its
+# expression. Only
 # `average_expression_all_cells` reaches this table through the per-simulation output, so every such
 # analysis has had to load a 16 MB sim_input.rds to find the other half. Joining it here costs one
 # read and makes the summary self-sufficient.