feat: completude de l'editeur Excel - mise en forme, sortie, concurrence, cache, outils IA #156

L'editeur lisait les styles mais ne les ecrivait pas, exportait la feuille
entiere et laissait le dernier-ecrivain gagner entre processus.

- A8 : bouton Mise en forme (gras/italique/souligne, alignements, couleurs via
  selecteur natif, formats de nombre, fusion, volets figes, largeur/hauteur) et
  nouvelle route PUT .../xlsx/style (verrou, backup, swap atomique, garde de
  perte, If-Match) ; A9 : decision "pas de moteur de formule" annoncee dans
  l'UI ; A10 : undo/redo unifie, piles par fichier conservees au re-rendu
- A11 : export de la selection + Markdown/HTML/impression et recherche sur
  toutes les feuilles ; A12 : concurrence optimiste (If-Match -> 409 reparable,
  retry qui relit) ; A13 : cache des metadonnees par (chemin, mtime, taille)
- A14 : outils IA .xlsm/.csv + search_workbook, analyze_range,
  edit_xlsx_structure
- tests : test_xlsx_styles.py (17), test_spreadsheet_tools.py (34),
  TestOptimisticConcurrency/TestMetaCache, JSDOM xlsx-viewer 107/107

backend/xlsx_reader.py

@@ -14,9 +14,11 @@ read-only; :func:`render_csv_table` turns a CSV into the same table shape.
 
 from __future__ import annotations
 
+import csv
 import html
 import logging
 import re
+import threading
 import zipfile
 from datetime import date, datetime
 from pathlib import Path
@@ -65,8 +67,22 @@ LOSSY_PARTS: dict[str, tuple[str, ...]] = {
 # Excel recalculates.
 _CACHED_FORMULA_RE = re.compile(rb"<f[ >][^<]*</f>\s*<v>[^<]")
 
-# Sheet XML scanned by the cached-formula probe (CPU guard, like MAX_REPLACE_FILE_BYTES).
-_MAX_PROBE_BYTES = 8_000_000
+# Sheet XML scanned by the cached-formula probe (CPU guard, like
+# MAX_REPLACE_FILE_BYTES). BUG-099 — the allowance is PER SHEET: a single huge
+# first sheet used to eat the whole budget and hide a cached formula sitting in
+# the next one. The total ceiling still bounds the work on a many-sheet archive.
+# Running out of budget is reported as *unverified* (never as "nothing to lose")
+# so the write guard stays cautious instead of silently dropping the values.
+_MAX_PROBE_BYTES_PER_SHEET = 4_000_000
+_MAX_PROBE_BYTES_TOTAL = 32_000_000
+
+# #156-A13 — metadata cache. `read_sheet_window()` serves one window at a time
+# and used to reload the whole workbook (normal mode, data_only=False) for every
+# window, just to read the style/merge/freeze maps of one sheet. The result is
+# keyed by (path, mtime_ns, size): any write replaces the file, hence the key.
+_META_CACHE_MAX = 8
+_meta_cache: dict[str, tuple[tuple[int, int], dict[str, dict[str, Any]]]] = {}
+_meta_cache_lock = threading.Lock()
 
 # #153 A17 — OPC parts of chart / pivot objects, matched against the archive
 # name list (xl/charts/chart1.xml, xl/pivotTables/pivotTable1.xml, …).
@@ -111,6 +127,10 @@ def _cell_fragments(cell: Any) -> tuple[list[str], str | None]:
         fragments.append("font-weight:600")
     if font and font.italic:
         fragments.append("font-style:italic")
+    # #156-A8 — underline is written from the viewer too, so it is read back
+    # (the toggle in the formatting menu needs to see its own effect).
+    if font and font.underline:
+        fragments.append("text-decoration:underline")
     fmt = cell.number_format
     if fmt and fmt not in ("General", "@"):
         # A custom number format is signalled typographically (mono font)
@@ -320,24 +340,65 @@ def _table(
     return "".join(out)
 
 
-def _has_cached_formulas(zf: zipfile.ZipFile) -> bool:
-    """True when at least one formula cell still carries its computed value."""
-    budget = _MAX_PROBE_BYTES
+def _is_sheet_xml(name: str) -> bool:
+    """True for the worksheet XML parts the cached-formula probe scans."""
+    return name.startswith("xl/worksheets/sheet") and name.endswith(".xml")
+
+
+def _scan_cached_formulas(zf: zipfile.ZipFile) -> tuple[bool, bool]:
+    """Scan the sheet XML for a formula carrying a non-empty cached result.
+
+    Returns ``(found, unverified)``. ``unverified`` is True when the byte
+    budget stopped the scan before every sheet could be read to the end: a
+    negative result is then **not** proof that the workbook holds no cached
+    value (BUG-099), so callers must not treat it as a licence to write.
+
+    Each sheet gets its own :data:`_MAX_PROBE_BYTES_PER_SHEET` allowance (a
+    single huge sheet can no longer starve the others) while
+    :data:`_MAX_PROBE_BYTES_TOTAL` bounds the whole archive. A hit short-
+    circuits the scan: the answer is already known.
+    """
+    budget = _MAX_PROBE_BYTES_TOTAL
+    unverified = False
     for name in zf.namelist():
-        if not name.startswith("xl/worksheets/sheet") or not name.endswith(".xml"):
+        if not _is_sheet_xml(name):
             continue
+        sheet_budget = min(_MAX_PROBE_BYTES_PER_SHEET, budget)
+        exhausted = False
         try:
             with zf.open(name) as fh:
-                while budget > 0:
-                    chunk = fh.read(65536)
+                while sheet_budget > 0:
+                    # Read at most what the sheet's allowance has left, so one
+                    # large chunk can never consume the whole total budget.
+                    chunk = fh.read(min(65536, sheet_budget))
                     if not chunk:
-                        break
+                        break  # read to the end: this sheet is verified clean
+                    sheet_budget -= len(chunk)
                     budget -= len(chunk)
                     if _CACHED_FORMULA_RE.search(chunk):
-                        return True
+                        return True, False
+                else:
+                    # Left the loop on the budget, not on EOF.
+                    exhausted = True
         except (KeyError, OSError, zipfile.BadZipFile):
             continue
-    return False
+        if exhausted:
+            unverified = True
+        if budget <= 0:
+            unverified = True
+            break
+    return False, unverified
+
+
+def _has_cached_formulas(zf: zipfile.ZipFile) -> bool:
+    """True when at least one formula cell still carries its computed value.
+
+    Boolean view of :func:`_scan_cached_formulas` for the display path (the
+    second ``data_only=True`` read): a truncated scan simply skips the shadow
+    grid, it never claims the workbook is lossless.
+    """
+    found, _ = _scan_cached_formulas(zf)
+    return found
 
 
 def inspect_workbook(file_path: Path) -> list[str]:
@@ -349,6 +410,10 @@ def inspect_workbook(file_path: Path) -> list[str]:
 
     ``cached_values`` is a synthetic key: openpyxl keeps the formula but drops
     the cached result, so the workbook stays correct once Excel recalculates it.
+    ``cached_values_unverified`` (BUG-099) is the other synthetic key: the
+    cached-value probe ran out of budget, so a negative result is not proof —
+    the entry keeps the write guard cautious (409 + confirmation) rather than
+    promising a lossless round-trip it cannot vouch for.
     """
     try:
         with zipfile.ZipFile(file_path) as zf:
@@ -358,13 +423,27 @@ def inspect_workbook(file_path: Path) -> list[str]:
                 for key, prefixes in LOSSY_PARTS.items()
                 if any(name.startswith(prefix) for name in names for prefix in prefixes)
             }
-            if _has_cached_formulas(zf):
+            cached_found, cached_unverified = _scan_cached_formulas(zf)
+            if cached_found:
                 found.add("cached_values")
+            elif cached_unverified:
+                found.add("cached_values_unverified")
             return sorted(found)
     except (OSError, zipfile.BadZipFile):
         return []
 
 
+def invalidate_workbook_meta(file_path: Path | str) -> None:
+    """Drop the cached metadata of one workbook (#156-A13).
+
+    Called by the write path right after the atomic replace: the (mtime, size)
+    key already changes on a rewrite, this only closes the theoretical window
+    where a same-size write lands on the same timestamp tick.
+    """
+    with _meta_cache_lock:
+        _meta_cache.pop(str(file_path), None)
+
+
 def read_workbook_meta(file_path: Path) -> dict[str, dict[str, Any]]:
     """Return ``{sheet: {styles, aligns, merges, freeze}}`` for every sheet.
 
@@ -372,7 +451,31 @@ def read_workbook_meta(file_path: Path) -> dict[str, dict[str, Any]]:
     fragments are the workbook's own values, a failure yields ``{}`` per sheet
     so the viewer keeps its plain rendering. Styles are read with
     ``data_only=False`` — the edited value is the formula, not its result.
+
+    #156-A13 — the result is cached on ``(path, mtime_ns, size)``: a truncated
+    sheet is fetched window by window, and each window used to pay a full
+    workbook load for these maps alone. Callers only read from the mapping.
     """
+    try:
+        st = file_path.stat()
+    except OSError:
+        return {}
+    key = str(file_path)
+    stamp = (st.st_mtime_ns, st.st_size)
+    with _meta_cache_lock:
+        hit = _meta_cache.get(key)
+        if hit and hit[0] == stamp:
+            return hit[1]
+    out = _read_workbook_meta_uncached(file_path)
+    with _meta_cache_lock:
+        _meta_cache[key] = (stamp, out)
+        while len(_meta_cache) > _META_CACHE_MAX:
+            _meta_cache.pop(next(iter(_meta_cache)))
+    return out
+
+
+def _read_workbook_meta_uncached(file_path: Path) -> dict[str, dict[str, Any]]:
+    """Load the workbook once and build the per-sheet metadata maps."""
     try:
         wb = load_workbook(str(file_path), data_only=False)
     except Exception:
@@ -781,17 +884,46 @@ def read_workbook_dashboard(file_path: Path) -> dict[str, Any]:
 # ── #153 A16 — additional spreadsheet formats ───────────────────────────────
 
 
-def render_csv_table(raw: str, *, delimiter: str = ",") -> str:
+# BUG-098 — a CSV written by a French office suite is `;`-separated, and the
+# delimiter must be *detected* (then reused on write-back), not assumed.
+CSV_SNIFF_BYTES = 4096
+_CSV_DELIMITERS = (",", ";", "\t", "|")
+
+
+def sniff_csv_delimiter(raw: str) -> str:
+    """Return the most likely field delimiter of *raw* (``,`` as fallback).
+
+    :class:`csv.Sniffer` handles quoted fields and multi-line values; it is
+    unreliable on short or single-column samples, so the first non-empty line
+    is counted as a tie-breaker and a comma remains the last resort.
+    """
+    sample = raw[:CSV_SNIFF_BYTES]
+    try:
+        return csv.Sniffer().sniff(sample, delimiters="".join(_CSV_DELIMITERS)).delimiter
+    except csv.Error:
+        pass
+    first = next((line for line in sample.splitlines() if line.strip()), "")
+    counts = {delim: first.count(delim) for delim in _CSV_DELIMITERS}
+    best = max(counts, key=lambda delim: counts[delim])
+    return best if counts[best] else ","
+
+
+def render_csv_table(raw: str, *, delimiter: str | None = None) -> str:
     """Render CSV text as the same HTML table shape the xlsx viewer consumes.
 
     Row numbers replace the A1 column: a CSV has no fixed column count, so
     the first row is a plain data row like the others (the viewer offers the
     toolbar either way). Every cell is HTML-escaped at render time.
+
+    ``delimiter`` defaults to the sniffed one (:func:`sniff_csv_delimiter`,
+    BUG-098): a `;`-separated file used to render as a single column.
     """
     import csv as csv_mod
     import io as io_mod
 
-    reader = csv_mod.reader(io_mod.StringIO(raw), delimiter=delimiter)
+    reader = csv_mod.reader(
+        io_mod.StringIO(raw), delimiter=delimiter or sniff_csv_delimiter(raw)
+    )
     try:
         rows = [row for row in reader]
     except csv_mod.Error: