#!/usr/bin/env Rscript
# analyse_stadium_relaunch.R
# StarForth LithosAnanke — EXEC-DOE 3x3 Latin square campaign,
# re-run on the post-item-4.6 Stadium substrate (2026-08-20).
#
# Generates figures + tables for stadium_relaunch_report.tex
#
# Data source: runs/acl-rwt-20260820/{arch}-seed{seed}.csv
#   9 cells: {amd64,aarch64,riscv64} x {12345,67890,13579}
#   480 rows/cell (16 L8 configs x 30 reps, Fisher-Yates shuffled)
#
# Scope note: this campaign validates that the EXEC-DOE mechanism runs
# cleanly and completely on the Stadium substrate (item 5.1's own concern:
# "a green POST suite is not evidence that determinism holds"). It is NOT
# an ACL-RWT overhead measurement -- ACL.4th is not self-activated in this
# repo's default init.4th, so these cells ran with ACL inactive.

suppressPackageStartupMessages({
  library(ggplot2)
  library(svglite)
  library(dplyr)
  library(tidyr)
  library(scales)
  library(patchwork)
})

SCRIPT_DIR <- tryCatch(
  dirname(normalizePath(sys.frames()[[1]]$ofile)),
  error = function(e) getwd()
)
BASE_DIR   <- normalizePath(file.path(SCRIPT_DIR, ".."))
RUNS_DIR   <- file.path(BASE_DIR, "runs", "acl-rwt-20260820")
OUT_CHARTS <- file.path(SCRIPT_DIR, "charts")
OUT_TABLES <- file.path(SCRIPT_DIR, "tables")
dir.create(OUT_CHARTS, showWarnings = FALSE, recursive = TRUE)
dir.create(OUT_TABLES, showWarnings = FALSE, recursive = TRUE)

cat("══════════════════════════════════════════════════════════════════\n")
cat("  StarForth LithosAnanke — Stadium-substrate EXEC-DOE relaunch\n")
cat("  Item 5.1 / F.3 — campaign-mechanism validation, 2026-08-20\n")
cat("══════════════════════════════════════════════════════════════════\n\n")

# ── palette (matches analyse_acl_rwt.R) ───────────────────────────────────────
arch_colours <- c(amd64   = "#E07B39",
                  aarch64 = "#4A90D9",
                  riscv64 = "#50C878")

theme_light_sf <- function(base = 11) {
  theme_minimal(base_size = base) %+replace% theme(
    panel.grid.minor  = element_blank(),
    panel.grid.major  = element_line(colour = "grey90"),
    strip.text        = element_text(face = "bold"),
    plot.title        = element_text(face = "bold", size = base + 1),
    plot.subtitle     = element_text(colour = "grey40", size = base - 2),
    legend.position   = "bottom",
    legend.key.size   = unit(0.5, "cm")
  )
}

theme_dark_sf <- function(base = 11) {
  theme_minimal(base_size = base) %+replace% theme(
    panel.background  = element_rect(fill = "#0d0d0d", colour = NA),
    plot.background   = element_rect(fill = "#0d0d0d", colour = NA),
    panel.grid.major  = element_line(colour = "#1e1e1e"),
    panel.grid.minor  = element_blank(),
    axis.text         = element_text(colour = "#aaaaaa"),
    axis.title        = element_text(colour = "#cccccc"),
    strip.text        = element_text(colour = "white", face = "bold"),
    plot.title        = element_text(colour = "white", face = "bold", size = base + 1),
    plot.subtitle     = element_text(colour = "#666666", size = base - 2),
    legend.text       = element_text(colour = "#aaaaaa"),
    legend.title      = element_text(colour = "#cccccc"),
    legend.background = element_rect(fill = "#0d0d0d", colour = NA),
    legend.position   = "bottom",
    legend.key.size   = unit(0.5, "cm")
  )
}

save_svg <- function(plot, name, w = 12, h = 7) {
  path <- file.path(OUT_CHARTS, paste0(name, ".svg"))
  svglite(path, width = w, height = h)
  print(plot)
  dev.off()
  cat(sprintf("  Saved: %s.svg\n", name))
  invisible(path)
}

# ── load data ──────────────────────────────────────────────────────────────
col_names <- c("run_id", "cfg", "rep", "ent_in", "cv_in", "tmp_in", "stb_in",
               "l8_mode", "win_div", "infer_win", "infer_dec_q", "infer_var_q",
               "early_exit", "bc_mean_q", "bb_mean_q", "fit_q")

archs <- c("amd64", "aarch64", "riscv64")
seeds <- c("12345", "67890", "13579")

load_cell <- function(arch, seed) {
  f <- file.path(RUNS_DIR, sprintf("%s-seed%s.csv", arch, seed))
  raw <- readLines(f)
  data_lines <- raw[grepl("^[0-9]", raw)]
  df <- read.csv(text = paste(data_lines, collapse = "\n"),
                  header = FALSE, col.names = col_names,
                  strip.white = TRUE)
  df$arch <- arch
  df$seed <- seed
  df
}

cat("Loading 9 cells...\n")
all_data <- bind_rows(lapply(archs, function(a) {
  bind_rows(lapply(seeds, function(s) load_cell(a, s)))
}))
cat(sprintf("  Total rows loaded: %d (expect 4320)\n", nrow(all_data)))

# ── per-cell completeness table ───────────────────────────────────────────
cell_summary <- all_data %>%
  group_by(arch, seed) %>%
  summarise(
    n_rows      = n(),
    n_distinct_run_id = n_distinct(run_id),
    l8_modes    = n_distinct(l8_mode),
    mean_win_div = mean(win_div),
    mean_fit_q   = mean(fit_q),
    mean_bc_mean_q = mean(bc_mean_q),
    mean_bb_mean_q = mean(bb_mean_q),
    .groups = "drop"
  )
cell_summary$arch <- factor(cell_summary$arch, levels = archs)
write.csv(cell_summary, file.path(OUT_TABLES, "stadium_relaunch_cell_summary.csv"),
          row.names = FALSE)
cat("\nCell summary:\n")
print(as.data.frame(cell_summary))

# ── cross-arch / cross-seed invariance tests (Kruskal-Wallis, matches the
#    original report's own non-parametric methodology for this kind of
#    campaign-level distributional comparison) ─────────────────────────────
kw_arch_fit  <- kruskal.test(fit_q ~ arch, data = all_data)
kw_arch_win  <- kruskal.test(win_div ~ arch, data = all_data)
kw_seed_fit  <- kruskal.test(fit_q ~ seed, data = all_data)
kw_seed_win  <- kruskal.test(win_div ~ seed, data = all_data)

kw_results <- capture.output({
  cat("Kruskal-Wallis: fit_q ~ arch\n"); print(kw_arch_fit); cat("\n")
  cat("Kruskal-Wallis: win_div ~ arch\n"); print(kw_arch_win); cat("\n")
  cat("Kruskal-Wallis: fit_q ~ seed\n"); print(kw_seed_fit); cat("\n")
  cat("Kruskal-Wallis: win_div ~ seed\n"); print(kw_seed_win); cat("\n")
})
writeLines(kw_results, file.path(OUT_TABLES, "stadium_relaunch_kw_results.txt"))
cat("\n")
cat(paste(kw_results, collapse = "\n"))
cat("\n\n")

# ══════════════════════════════════════════════════════════════════════════
# FIGURE 1: 3x3 Latin square completeness heatmap (rows captured per cell)
# ══════════════════════════════════════════════════════════════════════════
cat("[SR-1] Completeness heatmap (light + dark)...\n")

df_heat <- cell_summary %>% mutate(seed = factor(seed, levels = seeds))

make_heatmap <- function(dark = FALSE) {
  thm <- if (dark) theme_dark_sf() else theme_light_sf()
  txt <- if (dark) "white" else "grey10"
  ggplot(df_heat, aes(x = seed, y = arch, fill = n_rows)) +
    geom_tile(colour = if (dark) "#0d0d0d" else "white", linewidth = 1.5) +
    geom_text(aes(label = sprintf("%d/480", n_rows)), colour = txt,
              fontface = "bold", size = 4.2) +
    scale_fill_gradient(low = "#c0392b", high = "#2CA02C", limits = c(0, 480),
                         name = "Rows captured") +
    labs(
      title = "EXEC-DOE Campaign Completeness — Stadium Substrate",
      subtitle = "3x3 Latin square: 3 seeds x 3 ISAs, 30 reps x 16 L8 configs per cell (480 rows/cell)",
      x = "Seed", y = "Architecture"
    ) +
    thm
}
save_svg(make_heatmap(FALSE), "stadium_relaunch_completeness_light", w = 9, h = 6)
save_svg(make_heatmap(TRUE),  "stadium_relaunch_completeness_dark",  w = 9, h = 6)

# ══════════════════════════════════════════════════════════════════════════
# FIGURE 2: fit_q distribution by architecture (all seeds pooled)
# ══════════════════════════════════════════════════════════════════════════
cat("[SR-2] fit_q distribution by architecture (light + dark)...\n")

all_data$arch <- factor(all_data$arch, levels = archs)

make_fit_box <- function(dark = FALSE) {
  thm <- if (dark) theme_dark_sf() else theme_light_sf()
  ggplot(all_data, aes(x = arch, y = fit_q, fill = arch)) +
    geom_boxplot(alpha = 0.85, outlier.size = 0.6, outlier.alpha = 0.4) +
    scale_fill_manual(values = arch_colours, guide = "none") +
    scale_x_discrete(labels = c(amd64 = "amd64\n(x86-64)",
                                 aarch64 = "aarch64\n(ARMv8-A)",
                                 riscv64 = "riscv64\n(RV64GC)")) +
    labs(
      title = "Inference-Fit Distribution Across ISAs",
      subtitle = sprintf(
        "Kruskal-Wallis fit_q ~ arch: H=%.3f, p=%.4f (all 4,320 rows, 3 seeds pooled per ISA)",
        kw_arch_fit$statistic, kw_arch_fit$p.value
      ),
      x = "Instruction-Set Architecture", y = "fit_q (Q16 fixed-point)"
    ) +
    thm
}
save_svg(make_fit_box(FALSE), "stadium_relaunch_fitq_box_light", w = 9, h = 6)
save_svg(make_fit_box(TRUE),  "stadium_relaunch_fitq_box_dark",  w = 9, h = 6)

# ══════════════════════════════════════════════════════════════════════════
# FIGURE 3: mean win_div per (arch, seed) cell — grouped bar
# ══════════════════════════════════════════════════════════════════════════
cat("[SR-3] Mean window-diversity per cell (light + dark)...\n")

df_bar <- cell_summary %>% mutate(seed = factor(seed, levels = seeds))

make_windiv_bar <- function(dark = FALSE) {
  thm <- if (dark) theme_dark_sf() else theme_light_sf()
  ggplot(df_bar, aes(x = seed, y = mean_win_div, fill = arch)) +
    geom_col(position = position_dodge(width = 0.75), width = 0.65,
             colour = NA, alpha = 0.92) +
    scale_fill_manual(values = arch_colours, name = "ISA") +
    labs(
      title = "Mean Window-Diversity per Campaign Cell",
      subtitle = "9 cells, 480 runs each, Stadium substrate (post item-4.6 quota-grant fix)",
      x = "Seed", y = "Mean win_div"
    ) +
    thm
}
save_svg(make_windiv_bar(FALSE), "stadium_relaunch_windiv_bar_light", w = 9, h = 6)
save_svg(make_windiv_bar(TRUE),  "stadium_relaunch_windiv_bar_dark",  w = 9, h = 6)

cat("\nDone. Tables in analysis/tables/, charts in analysis/charts/.\n")