fix(knesset_committee_decisions): paginate OData responses

[wilfoa]

Summary

pipelines/knesset/knesset_committee_decisions.py calls
KNS_DocumentCommitteeSession once at the top of flow() and ignores
@odata.nextLink. OData v4 on knesset.gov.il paginates at 100 rows,
so the pipeline only ever sees the first 100 documents — sorted by
Id ascending those are all from 2016 (Knesset 18/20). The output
index.csv has been frozen at 7 records from 2016 since.

Live API today reports 12.5M+ matching documents and the most recent
GroupTypeID=106 PDF is dated 2026-04-21. The data is there; the
pipeline just isn't reading it.

What this PR does

• Adds an _odata_paged() helper that walks every @odata.nextLink.
• Replaces the global KNS_DocumentCommitteeSession fetch with a
  per-session paginated query (CommitteeSessionID eq … and GroupTypeID eq 106). Each per-session response is small (typically
  0–3 PDFs) and bounded.
• Also paginates KNS_Committee and KNS_CommitteeSession for
  consistency — the current dataset's committee/session counts are
  small enough today that they fit in 100 rows, but this prevents the
  same silent regression next year.
• Renames the inner document loop variable so it no longer shadows
  the outer dict (the previous code reassigned document = requests.get(...) mid-loop, which would have broken the next
  iteration's document['ApplicationDesc'] access if that path had
  been reached more than once for the same dict).

Smoke-test against live OData (no dpp run needed)

# Per-session filter works:
session=2242368 Id=12594486 date=2026-04-21T09:06:13.71+03:00
session=2201184 Id=12577930 date=2026-04-20T11:04:11.54+03:00
session=2201184 Id=12577921 date=2026-04-20T11:02:13.193+03:00

# Pagination walks correctly:
committee 2024: walked 62 sessions across multiple pages

The first-page-only behavior of the existing code is reproducible by
just curling the OData URL — $top defaults to 100, response includes
@odata.nextLink, only documents 334231–341207 (all 2016) appear in
that first page when $orderby=Id ascending.

Test plan

• Live OData smoke (above) confirming the pagination + per-session
  filter work and find current data.
• Local dpp run and inspect that index.csv has thousands of
  rows spanning 2016 → 2026 instead of 7 from 2016.
• After merge + pipeline rebuild, confirm
  next.obudget.org/datapackages/knesset/knesset_committee_decisions/datapackage.json
  reports count_of_rows > 7.

Why this matters downstream

whiletrue-industries/rebuilding-bots consumes
next.obudget.org/datapackages/knesset/knesset_committee_decisions/index.csv
to power one of the unified Parlibot's contexts. With the bug, the bot
has been answering committee-decision questions from a 7-row, 2016-only
corpus for the past several months. We just shipped a Layer-1 safety
rail there (refuses to overwrite local extraction CSVs when upstream
is empty / stale, alarms via SNS). This PR is Layer 2 — the actual
upstream fix.

🤖 Generated with Claude Code