#!/usr/bin/env python
"""Stage 16 — assemble a single self-contained HTML report of the v2 results for review.

Embeds every figure as base64 (so the .html is portable) and renders the score tables. Covers:
  * headline model x block binary skill (all families x cutoffs)
  * EHD_ens: ensemble (mean + LOYO stack) vs zoo members, table + figure
  * per-cell skill maps per lead week (POD/POFD/FAR/CSI/EDI/SEDI) for ens + members
Run stages 13, 14, 15 first. Output: results/v2_report.html
"""
from __future__ import annotations
import os, sys, base64
import pandas as pd

sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "src"))

HERE = os.path.dirname(os.path.abspath(__file__))
ROOT = os.path.dirname(HERE)
TABLES = os.path.join(ROOT, "results", "tables")
FIGS = os.path.join(ROOT, "results", "figures")
OUT = os.path.join(ROOT, "results", "v2_report.html")


def img(path, width="100%"):
    if not os.path.exists(path):
        return f'<p style="color:#b00">[missing figure: {os.path.basename(path)}]</p>'
    b = base64.b64encode(open(path, "rb").read()).decode()
    return f'<img style="width:{width};max-width:1400px;border:1px solid #ddd" ' \
           f'src="data:image/png;base64,{b}"/>'


def table(path, highlight_ens=False, cols=None):
    if not os.path.exists(path):
        return f'<p style="color:#b00">[missing table: {os.path.basename(path)}]</p>'
    df = pd.read_csv(path)
    if cols:
        df = df[[c for c in cols if c in df.columns]]
    def rowclass(r):
        if highlight_ens and str(r.get("model", "")).startswith("ens_"):
            return ' class="ens"'
        return ""
    head = "".join(f"<th>{c}</th>" for c in df.columns)
    body = ""
    for _, r in df.iterrows():
        body += f"<tr{rowclass(r)}>" + "".join(f"<td>{r[c]}</td>" for c in df.columns) + "</tr>"
    return f'<table><thead><tr>{head}</tr></thead><tbody>{body}</tbody></table>'


CSS = """
body{font-family:-apple-system,Segoe UI,Roboto,Helvetica,Arial,sans-serif;margin:0;
     color:#1b1b1b;line-height:1.5}
header{background:#0d3b66;color:#fff;padding:24px 40px}
header h1{margin:0 0 4px 0;font-size:24px}
header p{margin:0;opacity:.85}
main{max-width:1500px;margin:0 auto;padding:0 40px 80px}
h2{border-bottom:3px solid #0d3b66;padding-bottom:6px;margin-top:48px;color:#0d3b66}
h3{margin-top:28px;color:#144}
table{border-collapse:collapse;font-size:12.5px;margin:12px 0;width:100%}
th,td{border:1px solid #ccc;padding:4px 8px;text-align:right}
th{background:#eef3f8;position:sticky;top:0}
td:first-child,th:first-child{text-align:left}
tr.ens{background:#fff3cd;font-weight:600}
.note{background:#f4f7fb;border-left:4px solid #0d3b66;padding:10px 16px;margin:16px 0;
      font-size:14px}
.warn{background:#fff4e5;border-left:4px solid #e08600}
figure{margin:16px 0}
figcaption{font-size:13px;color:#555;margin-top:6px}
nav{position:sticky;top:0;background:#fff;border-bottom:1px solid #ddd;padding:8px 40px;
    font-size:13px;z-index:10}
nav a{margin-right:16px;color:#0d3b66;text-decoration:none}
code{background:#eee;padding:1px 5px;border-radius:3px}
"""


def main():
    from datetime import date
    import glob
    has_torch = bool(glob.glob(os.path.join(ROOT, "outputs", "loyo_pred_v2_cnn",
                                            "pred_v2cnn_*.npz")))
    parts = []
    parts.append(f"<header><h1>S2S Extreme-Heat-Day (EHDE) forecast — v2 results</h1>"
                 f"<p>Vietnam · severity predictand + model zoo + cascade + <b>EHD_ens</b> · "
                 f"20-fold leave-one-year-out, pooled out-of-sample 2003–2022 · "
                 f"generated {date.today().isoformat()}</p></header>")
    parts.append('<nav><a href="#overview">Overview</a>'
                 '<a href="#ensemble">EHD_ens</a>'
                 '<a href="#zoo">Model zoo</a>'
                 '<a href="#maps">Skill maps</a>'
                 '<a href="#vote">Vote ensemble &amp; regions</a>'
                 '<a href="#methods">Methods</a></nav>')
    parts.append("<main>")

    parts.append('<h2 id="overview">Overview</h2>')
    parts.append('<div class="note">All scores are <b>out-of-sample</b>: each year was held out, '
                 'everything (threshold, models, recalibration, decision thresholds, and the '
                 'ensemble stacker) was fit without it, and the 20 held-out years were pooled '
                 '(leakage rule). Event = <b>EHDE</b> = ≥1 EHD in the lead block; '
                 'EHD day = Tx ≥ min(Tx90p, 35°C). Six fixed lead blocks W1–W6 (0–10 … 39–45 d).'
                 '</div>')
    if has_torch:
        parts.append('<div class="note">Model set: 7 sklearn members + <b>CNN</b> '
                     '(fully-convolutional, spatial) + <b>deep ANN</b> (torch), all folded into '
                     'EHD_ens.</div>')
    else:
        parts.append('<div class="note warn">CNN and a deeper torch ANN are <b>queued</b> (need a '
                     'gridded input pipeline + a fresh SLURM re-run); this report covers the 7 '
                     'tabular members and their ensembles. It will be regenerated once CNN/ANN '
                     'land.</div>')

    parts.append('<h2 id="ensemble">EHD_ens — ensemble of AI members</h2>')
    parts.append('<p><b>ens_mean</b> = unweighted mean of the 7 member probabilities. '
                 '<b>ens_stack</b> = logistic meta-learner over the members, trained '
                 'leave-one-year-out at the meta level (leakage-safe).</p>')
    parts.append(f'<figure>{img(os.path.join(FIGS, "v2_ensemble_skill_by_lead.png"))}'
                 '<figcaption>EHD_ens (bold) vs the 7 zoo members (grey): SEDI and Brier skill '
                 'by lead block.</figcaption></figure>')
    parts.append('<h3>Ensemble vs members — all binary scores (cutoff 1.0)</h3>')
    parts.append(table(os.path.join(TABLES, "v2_scores_with_ensemble.csv"), highlight_ens=True))

    parts.append('<h2 id="zoo">Model zoo — binary skill by lead block</h2>')
    parts.append(f'<figure>{img(os.path.join(FIGS, "v2_binary_skill_by_lead.png"))}'
                 '<figcaption>Per-family SEDI and Peirce by lead (each family at its best '
                 'cascade cutoff).</figcaption></figure>')
    parts.append('<h3>Best model per block (by SEDI)</h3>')
    parts.append(table(os.path.join(TABLES, "v2_best_model_by_block.csv")))
    parts.append('<h3>Full table — every family × cascade cutoff × block</h3>')
    parts.append(table(os.path.join(TABLES, "v2_cv_scores_by_model_block.csv")))

    parts.append('<h2 id="maps">Per-cell skill maps by lead week</h2>')
    parts.append('<p>Each panel is a grid-cell map; rows are metrics '
                 '(POD, POFD, FAR, CSI, EDI, SEDI), columns are lead weeks W1–W6. '
                 'For POD/CSI/EDI/SEDI higher (yellow/red) = better; for POFD/FAR lower = '
                 'better. Full fields incl. HSS/Peirce/freq_bias/n are in '
                 '<code>outputs/v2_skill_maps.nc</code>.</p>')
    map_models = [("ens_stack", "EHD_ens (stacked) — the product"),
                  ("ens_mean", "EHD_ens (mean)"),
                  ("cnn", "CNN (fully-convolutional)"),
                  ("ann_torch", "Deep ANN (torch)"),
                  ("hgb", "HGB (single member)"),
                  ("mlp", "MLP (shallow sklearn member)")]
    for m, label in map_models:
        fig_path = os.path.join(FIGS, f"v2_skillmap_{m}.png")
        if not os.path.exists(fig_path):
            continue
        parts.append(f'<h3>{label}</h3>')
        parts.append(f'<figure>{img(fig_path)}</figure>')

    # ---------------- prompt section 6: vote-fraction ensemble + regional alerts -------------
    if os.path.exists(os.path.join(TABLES, "v6_vote_scores_by_block.csv")):
        parts.append('<h2 id="vote">Vote-fraction ensemble &amp; regional alerts (prompt §6)</h2>')
        parts.append('<div class="note">A second, model-agnostic ensemble: each of the '
                     '<b>23 AI members</b> (7 sklearn families × 3 cascade cutoffs + CNN + deep '
                     'ANN) casts a yEHDE vote using its own leakage-safe per-region threshold, and '
                     '<b>P(yEHDE)=m/n</b>. A deterministic forecast is EHDE=1 where P&gt;P₀, with '
                     'P₀ chosen to maximise Peirce (POD−POFD) leave-one-year-out.</div>')

        parts.append('<h3>§6.1 — Member comparison (POD high-good, POFD/FAR low-good)</h3>')
        parts.append(f'<figure>{img(os.path.join(FIGS, "v6_member_pod_pofd_far.png"))}'
                     '<figcaption>Per-member POD / POFD / FAR by lead block. CNN is the strongest '
                     'single member at long lead (W3–W6).</figcaption></figure>')
        parts.append('<p>Member contribution to the vote ensemble (leave-one-member-out ΔPeirce, '
                     'mean over W3–W6): all members are near-redundant (|ΔPeirce|≲0.002); CNN is '
                     'the top contributor. Full detail: <code>v6_member_contribution.csv</code>, '
                     '<code>v6_member_scores_by_block.csv</code>.</p>')

        parts.append('<h3>§6.2–6.3 — Vote fraction P(yEHDE)=m/n and optimal P₀</h3>')
        parts.append(f'<figure>{img(os.path.join(FIGS, "v6_p0_sweep.png"))}'
                     '<figcaption>Binary scores vs the vote threshold P₀ (pooled LOYO); dashed '
                     'line = Peirce-optimal P₀ (≈ a simple majority vote).</figcaption></figure>')
        parts.append('<p>Deterministic yEHDE at the per-fold Peirce-optimal P₀ '
                     '(leakage-safe headline):</p>')
        parts.append(table(os.path.join(TABLES, "v6_vote_scores_by_block.csv")))
        parts.append('<div class="note">The unweighted vote peaks near a <b>majority vote</b> '
                     '(P₀≈0.5) and roughly matches the best single member — a robust, '
                     'interpretable product, though the trained <b>ens_stack</b> above is a little '
                     'sharper at short lead.</div>')

        parts.append('<h3>§6.4 — Regional event forecast (M<sub>j</sub>/N<sub>j</sub>)</h3>')
        parts.append('<p>Per region R<sub>j</sub> and week W<sub>k</sub>, the forecast coverage '
                     'P=M<sub>j</sub>/N<sub>j</sub> (fraction of the region’s cells forecast '
                     'yEHDE) is compared with the same rule on the observed field; a region is '
                     'alerted if P≥P_thres. P_thres is chosen per week (Peirce, leakage-safe), '
                     'with a fixed-0.5 majority reference.</p>')
        parts.append(f'<figure>{img(os.path.join(FIGS, "v6_regional_event_heatmap.png"))}'
                     '<figcaption>Regional event-forecast Peirce (region × week). Aggregating to '
                     'region×week is markedly more skilful than per-cell.</figcaption></figure>')
        parts.append(table(os.path.join(TABLES, "v6_regional_by_week.csv")))
        parts.append('<div class="note warn">Per-fold Peirce selection of the regional P_thres is '
                     'jumpy across weeks (e.g. W5→0.10); the <code>*_at0p5</code> columns give the '
                     'steadier majority-vote reference. Per-region×week detail: '
                     '<code>v6_regional_event_table.csv</code>.</div>')

    parts.append('<h2 id="methods">Methods & provenance</h2>')
    parts.append('<ul>'
                 '<li><b>Members:</b> logistic, hgb, rf, mlp, ridge_sev, hgb_sev, mlp_sev '
                 '(single-member MOS; severity regressors predict Sev=Tx−Thr then isotonic→P).</li>'
                 '<li><b>Cascade:</b> LV1 easy-negative screen (cutoff sweep 0.90/0.99/1.0); '
                 'shown ~neutral, so ensembles use cutoff 1.0 members.</li>'
                 '<li><b>Ensemble threshold:</b> per-region Peirce-optimal, fit on the fit-years '
                 'ensemble probabilities, applied to the held-out year.</li>'
                 '<li><b>Maps:</b> per-cell 2×2 counts pooled over 20 held-out years, each sample '
                 'using its own fold/region decision threshold.</li>'
                 '<li><b>Tables:</b> results/tables/v2_*.csv · '
                 '<b>Fields:</b> outputs/v2_skill_maps.nc · '
                 '<b>Predictions:</b> outputs/loyo_pred_v2/, outputs/loyo_pred_v2_ens/.</li>'
                 '</ul>')
    parts.append("</main>")

    html = f"<!doctype html><html><head><meta charset='utf-8'>" \
           f"<title>v2 EHDE report</title><style>{CSS}</style></head>" \
           f"<body>{''.join(parts)}</body></html>"
    with open(OUT, "w") as f:
        f.write(html)
    print(f"[report] wrote {OUT} ({len(html)/1e6:.2f} MB)")


if __name__ == "__main__":
    main()