Land the ASBB P0 KB work, the SciLEx partial-collection fix, and green the CI

[lfnothias]

Land the ASBB P0 KB work, the SciLEx partial-collection fix, and green the CI

Consolidates every outstanding line of work into one releasable main. The repository is
already public, and its main has had a failing test job since 2026-07-10; this branch
closes that and brings the two stranded fix lines back together.

What it lands

Branch merges. origin/main (the zero-vector and session-database integrity fixes
#32/#33/#35/#36) and origin/asbb/p0-kb-doc (#34, plus the ASBB P0 KB work: symbol-level
code chunking, PMC-OA supplementary ingest, the supplementary endpoint). #34 had been
merged into a side branch only and was present on no other branch.

SciLEx partial collection. A single rate-limited backend (typically Semantic Scholar's
public tier) could outrun the aggregator's per-provider timeout, which cancelled the call
and discarded the results the fast backends had already collected — a query narrowed to
SciLEx sub-providers returned zero hits even when SciLEx had collected hundreds. The
adapter now takes a soft collection deadline, abandons stragglers, and returns partial
results with a structured partial_results_timeout warning.

Three defects found while hardening it, each with a regression test that fails without the
fix:

• The deadline is budgeted per call, not per pass. search() retries title-like
  queries with a normalized title; giving each pass the full deadline took up to 1.6× the
  caller's timeout, so the outer cancel fired and discarded both passes — reintroducing the
  exact failure the deadline exists to prevent.
• Abandoned backends keep writing into the scratch directory. Removing it inline raced them
  and raised Directory not empty, discarding collected results; suppressing that error
  instead leaked the directory permanently (measured: thousands of entries per abandoned
  search, unbounded in a long-running server). The directory is now reclaimed by a reaper
  once the abandoned backends actually stop.
• The partial-collection warning is carried in a call-local report rather than on the
  adapter. A worker thread can outlive a cancelled call, so adapter state leaked into the
  next search — a clean query could report another query's "partial results" warning, and
  the reverse interleaving hid a genuinely partial one.

WRRF citation attribution. select_wrrf_merged_documents merged each chunk with its
neighbours in the score-sorted list, which are not neighbours in the source document, and
stamped the current chunk's citation on the result. Merged context could therefore carry one
paper's text under another paper's citation. Merging is now restricted to chunks from the
same paper.

Provenance deletion. Deleting a conversation left its provenance rows and its
data/provenance/<id>.jsonl sidecar — which holds verbatim prompts and responses — in
place. Both are now purged, including a guard against a conversation id that would escape
the sidecar directory.

export_astra without the private stack. The tool hard-required the private,
unpublishable indicium package, so the repository's only ASTRA feature errored for every
public user; its output was also not a valid ASTRA record (Insights lacked the required
created_at and evidence). It now uses a small vendored projection, emits a valid
Analysis, and reports claims it could not ground rather than emitting them without
evidence. Its test file is no longer excluded from CI.

Why CI goes green

main's failure was six tests importing indicium transitively with no guard, after
1c5c6736 added a module-scope from indicium import ASB_BASE to
indicium_layer/queries.py. Those tests now skip cleanly when the private package is
absent, exactly as the repository's existing convention does elsewhere.

Four other failures on this branch were stale tests, not product bugs, and were fixed on the
test side: an embedding stub that was not awaitable once the symbol-wise path began
embedding in the orchestrator; a tree-sitter probe assertion that still named only the
legacy parser package; composite-adapter tests importing a private package unguarded; and
CLI ingest-mode tests that failed on any machine holding a populated local session database.

Verification

• Full unit suite: 2582 passed, 8 skipped, 0 failed.
• CI simulation — the private indicium stack hidden and ci.yml's ignore/deselect list
  applied: 2462 passed, 8 skipped, 17 deselected, 0 failed.
• uv lock resolves on a fresh clone (305 packages, no private dependency).
• Reviewed by an adversarial panel before opening; both blocking findings it raised are
  fixed above and covered by tests.

Not included

Splitting _scilex_search_sync (353 lines) and _collect_all_backends into smaller units
is a real cleanup but predates this work and belongs in its own change.