PPTX Follow-up and Visual Finalization

This is the complete normative Step 6 contract for vault-ingress. It covers post-batch PPTX selection, bounded extraction, current-schema admission, rendered-page evidence, matching, and cross-talk visual updates.

Runs once after all Step 3 batches have completed.

The bounded directory invocation walks every eligible deck and accepts no include list, so it is the extraction step that runs, unchanged. classify-pptx-evidence.py gates what happens to its results: persist a new receipt for exactly the records it reports needs_extraction, and leave every current record's receipt as it stands — rewriting one would replace a proven binding with an identical one and burn the proof for nothing. No additional predicate: unmatched entries and PDF-primary talks with a PPTX already report pending, so naming them separately can only overwrite a record the classifier proved current. The gate is never a boolean read: a stale, legacy, or unverifiable record can carry visual_extracted: true and still need a fresh receipt — see bootstrap-and-preflight.md for the command and its output contract. Use the bounded directory invocation from that same reference and select the root-relative results it names; do not replace it with **/*.pptx or a shell/per-file extraction loop. Reuse the exact config-derived {template_skip_arguments} and {directory_exclusion_arguments} (including zero arguments for either empty array), keep the explicit --directory flag, preserve every bounded skip receipt, and never admit a ~$ Office lock file. Require the public schema-v1 pptx_directory_batch envelope. Safely extracted results from a partial batch remain usable, but only complete: true authorizes a full-catalog conclusion, a claim that a missing deck does not exist, or absence inferred from an empty result. Legacy unversioned batch output has unknown completeness and must be rerun before any such claim. Require schema v4 for current analysis. Regenerate v0-v3 output and stop on an unknown future schema rather than interpreting missing fields as zero. When rendered pages were inspected, rerun that selected deck as one supervised single-artifact invocation with --rendered-pdf <path.pdf> and one or more --inspected-pages <PAGE|START-END> arguments so the extraction receipt binds the exact artifacts and covered pages. A non-empty archive_recovery blocks a required native-deck claim until the source is restored or re-exported. For every catalog finding or citation, OCR is affirmative only at the individual receipt level: use recovered_text only when that same receipt has trustworthy_text: true. Never promote ocr_text or an OCR channel's aggregate text directly; both retain low-confidence text for review.

PPTX matching rules: The .pptx files are in Conference/Year/TalkName.pptx and shownotes entries have conference and title fields. Fuzzy-match by: normalize conference names (strip year, "Days", "Conference"), match by date proximity and title substring. Skip Office locks beginning ~$, files with "static" in name, conflict copies matching (N).pptx, and files matching config.template_skip_patterns. Configured pptx_directory_exclusions prune a real directory only by case-insensitive exact component identity after symlink/reparse rejection; they do not match substrings or authored file names. Preserve the single pptx_batch_directory_excluded receipt and do not scan below it. Some talks have multiple .pptx files (one per delivery) — match to the closest date.

After 3+ extractions, populate slide-design-spec.md; after 5+, analyze cross-talk patterns (colors, fonts, footers).