eliziff gives PDF evidence a fixed address

The stage-3 first-query scan counted a line break sitting at span end
(one past the last word) as inside the span, so a phrase ending exactly
at end-of-line - any quote ending at a block boundary - was dropped on
the first query but found by the indexed path on the second. Match the
indexed path: only breaks strictly inside [start, end) count.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QUsMXzBsEJ27DjJVPSfW5o

prune() replaces ten conditional-spread pyramids with identical
persisted-JSON key order, attachmentFromStored() unifies the duplicate
stored-record validators, parsedState()/cachedPdf() dedupe the
resume-vs-read and downloaded-branch copies, and the lease layer
collapses SELECT+INSERT/CAS into one upsert whose semantics match the
original exactly (live lease refuses even the same owner; expired
lease taken over atomically under BEGIN IMMEDIATE). -301 lines.
mike-provider-pdf:v1 / v2 pointer / binding byte-shapes unchanged.

Verified: tsc clean; bridge suite 34/34 incl. lease-takeover and
cold-failure backoff, plus consumer suites green.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QUsMXzBsEJ27DjJVPSfW5o

Persistent parse cache keyed by file SHA-256 + parser name/version under
the app's local data home (parse-cache/), scoped per user/document so
matter content never crosses scopes. Wired into both production text
extraction paths (local extractLocalDocument and cloud
readDocumentContent) via a shared per-format parser table; content or
parser-version changes naturally miss, clearParseCache deletes a scope
or everything, hits log actually avoided work. Measured: 554KB PDF
1468ms -> 2ms, 121KB DOCX 128ms -> 1ms, byte-identical output.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yXxuitxX5h4wm1v88hnga

P0.8 remainder, local lane. localDocumentResponse hardcoded status:
"ready" for every document, so the Library could not tell a parsing,
degraded, or failed structural parse from a clean one without a
per-document round trip.

Backend: new peekLocalPdfParseState reads the durable job's state file
only - no artifact validation, no diagnostics load, no writes - so a
listing of N documents costs N stat-reads. localDocumentResponse
becomes async and carries parse_state (status, error, attempts,
timestamps, engine_status, cache_hit, page_count, diagnostic_count,
structural_repair_available) for PDF versions; null means no parse lane
(non-PDF / flat text). `status` keeps its storage-readiness meaning -
the parse lifecycle is a separate field, and flat text remains
available in every parse state.

Frontend: Document.parse_state typed; DocTable renders a lifecycle chip
beside the filename (spinner while queued/parsing, amber Degraded with
diagnostic count in the title, red Parse failed with the error in the
title and a Retry action). Retry posts the existing
actions/retry-pdf-parse route via the new retryLibraryPdfParse API and
refreshes the collection; clean ready parses and non-PDF rows render no
chip. LibraryWorkspace wires retryPdfParse per library kind.

Deliberately deferred, per the plan's own ordering ("cloud lane last"):
the cloud/Supabase lane still reports storage status only - its
documents carry no parse_state and the chip stays absent there.

Suites: localDocumentStore, localLibraryPdfParse, localPdfIngestion
(33), DocTable interaction/removal + LibraryWorkspace + FileDirectory
(29, incl. 3 new chip/retry tests); tsc clean both ends. types.ts
staged hunk-selectively - the unrelated in-flight assistant-event hunk
stays uncommitted.

Integration plan item 5 (task #28), local lane.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AgUm5EcRKbT3duVFMbrQD

Eli caught this on the test fixture: "no pages" on a PDF is a false statement
about the document.

Two things were wrong.

FALL-THROUGH. Both PDF routes end at the same engine. The ingested route
yields page BLOCKS via readLocalPdfSourceDoc; the un-ingested route runs
`legalpdf parse` and renders `[page N]` markers into the text it returns. The
map builder took the artifact when `parsed` existed and never tried the
markers, so a rendition that carried text but no page blocks was indexed as
pageless even though its own text was full of page markers. It now falls
through artifact -> markers.

VOCABULARY. "This document carries no page structure" conflated two states
that call for opposite responses. A DOCX genuinely has no fixed pagination
until something renders it. A PDF has pages whether or not we managed to read
them, so the same message there reports a gap in our pipeline as a fact about
the document -- exactly the class of error the refusal contract exists to
prevent. PageMap.source gains `unavailable`, set only where the file type is
known, and the two refusals now say different things: one tells you the
document is not paginated, the other tells you the index could not be built
and that page numbers appearing in the text are unverified.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H9ToHYJVDxfeJcJwdzrP2H