Built from the real-world workflow where forensic delay
        analysis starts from a folder containing schedule updates,
        owner correspondence, RFIs, change orders, and meeting
        minutes — all mixed together. The workbench produces:

          - Evidence ledger (chronological): all artifacts dated and
            summarized
          - Schedule chain-diff: 14-category manipulation log
            (TASKPRED add/remove, constraint flips, retroactive
            baseline edits, completion reversals)
          - Rolling baseline: per-activity baseline-at-introduction
            across the entire XER chain
          - Trust score: statistical impossibilities flagged
            (zero-duration-variance schedules, no-new-activities,
            every-activity-hits-baseline, etc.)
          - Slip-to-evidence cross-reference: each forensic slip
            auto-paired with documents in its window mentioning
            affected activity codes
          - Unified HTML dashboard with all of the above

        Use this tool when starting forensic delay analysis from raw
        evidence. For single-XER-pair forensic with hand-prepared
        events, use ``forensic_windows_analysis`` instead.

        Two input modes (supply exactly one):
          * ``folder_path`` — the ``staged_folder_path`` that POST /stage
            returned for YOUR upload, passed back exactly as received.
            No other folder is accepted: not a desktop path, not the
            server temp directory, not another folder inside it.
          * ``evidence_files`` — a CONTENT MANIFEST: a list of
            ``{"name": str, "content_b64": str}`` entries carrying
            base64-encoded file BYTES (handles binary PDF/XLSX/MSG as
            well as text). The tool decodes each blob, sanitizes the
            filename to a bare basename (rejecting path separators,
            ``..``, absolute/drive paths, control chars, dot-only
            traversal), writes it into a FRESH per-call tempdir under
            the allowed server-tempdir root, runs the analysis on that
            staged folder, then cleans the staged dir up. Caps: at most
            500 files and 60 MB total decoded bytes — an over-cap
            manifest returns a clear ``tool_error`` naming the cap and
            the actual size (NEVER silently truncated).

        Args:
            folder_path: the staged_folder_path from POST /stage
                (mode 1; must be that exact folder and must exist).
            evidence_files: content manifest (mode 2); list of
                ``{"name": str, "content_b64": str}``.
            output_dir: optional dir for outputs (tempdir if "").
            project_name: optional override.
            original_baseline_xer_filename: optional filename in the
                folder identifying the baseline XER.
            contract_form: contract template tag (default 'CCDC2').
            run_forensic: when True (default), also runs
                forensic_windows_analysis on the discovered XER chain.

        Returns:
            {
              "evidence_ledger":     {...},
              "chain_diff":          {...} | None,
              "rolling_baseline":    {...} | None,
              "trust_score":         {...} | None,
              "cross_reference":     {...} | None,
              "forensic_result":     {...} | None,
              "output_files":        {...},
              "errors":              {...} (per-step failure log)
            }