[spuentesp]

Landing the two pieces of pending loop work that were reviewed and are worth
keeping, rebased onto master (the original branch was 39 commits behind and
conflicted in three loop files).

Ingestion (2289ca7, cherry-picked unchanged — it applies cleanly)

• Capture the ingestion job id through an explicit on_job_created callback
  instead of monkey-patching mongodb_create_ingestion_job. The patch was
  module-level, so concurrent ingests could claim each other's job id.
• Extract each PDF page once per document instead of once per bookmark span,
  and keep the front matter before the first bookmark as its own section
  (those pages used to be dropped entirely).
• Classify a chunk from the deepest heading that carries a known tag, falling
  back through its ancestors.
• Record the section's page range in the chunk metadata, and add
  scripts/assess_ingestion_corpus.py to measure PDF coverage without an LLM.

Story arc (b30de84, a3e4d05)

A finished scene now advances the story instead of ending it, and the story
ends when its arc resolves.

The arc runs rising action → climax → falling action → resolution, and the
new scene_budget on StoryState (5 scenes by default) decides how long each
phase lasts. The old heuristic could never leave rising_action: tension
saturated below the climax threshold and no plot thread was ever closed, so
story_complete could never become true — the story had no way to end.

The story state is checkpointed after every scene, and start_or_resume()
returns that checkpoint instead of restarting the campaign, so an interrupted
campaign keeps its scenes, its arc position and its in-game time.

Counting a finished scene now happens in one place: the caller counts the
scene it just played and the transition only opens the next one. It used to
count it in both, so scenes_completed grew twice per scene.

Scene and conversation fixes (db68b96, b9cf7e9)

• finalize() closes the scene without resolving or narrating an empty turn.
• In a group conversation each NPC hears the replies the others already gave
  this round.

Verification

• packages/agents, packages/data-layer, packages/ui/backend and the
  repo-root tests/ suite
• ruff check, ruff format, check_layer_dependencies.py
• mypy on packages/agents/src — no new errors (the 12 that remain are
  pre-existing, in files this PR does not touch)
• New: packages/agents/tests/test_story_arc_contract.py, plus PDF page-bound
  tests in packages/data-layer/tests/test_db/test_ingest_tools.py

Mutation testing runs locally, not in CI. On the changed lines, every mutant
the code can produce is killed by the tests — 42 of 42, checked one at a time
(cosmic-ray needs longer than the budget on these modules).

Three pre-existing notes, all reproduced on master without these changes:

• typecheck fails on the strict-mypy backlog (the job is
  continue-on-error).
• unit-tests reaches 96% green and then hits the job's 20-minute timeout;
  master's run is cancelled the same way.
• mutation-diff exceeds the 600s per-module budget on indexer/agent.py,
  the same way every recent PR does.