A stored snapshot keeps no record of a partial export-table read

Index: Known gaps

Status (2026-11-03): FIXED for persistence and consumers; narrowed. ADR-079 is implemented, as its 2026-11-03 amendment specifies: the read is the binary section's abicheck.export_read extension fact (ADR-062 D19, no SCHEMA_VERSION bump), listing each platform block that was not read completely (partial, failed; {} for all complete), and must-understand for that section only when a value is partial or failed. A document without the fact reads unrecorded and keeps the older rule, never complete. The readers follow one effective-state rule (model.export_index.platform_export_read): a partial block withdraws its stored "not exported" answers at load, reads as a PARTIAL export fact (consumer scoping refuses it), makes the exports coverage record failed (export_table_partially_read), keeps each removal it cannot establish as COMPATIBLE_WITH_RISK under the partial_export_table_read rule, and lowers analysis_assurance.status to partial (with analysis_assurance.export_read naming the side, and a coverage row) and run_outcome.scope to incomplete. A live library read applies the same D1 rule, so live and stored operands agree. Regression tests: tests/unit/storage/test_export_read_persisted.py (property over the read outcomes, live and stored, ELF, PE and Mach-O; an older reader) and tests/test_cli_compare_partial_export_read.py (public compare, JSON and Markdown, exit under assurance.require_complete). Still open: (1) a baseline written before the fact whose read was cut short has no record and still reads by the older rule -- re-dump it; (2) in a release/package comparison a member's partial read lowers that member's assurance and annotates its removals, but does not mark the release's comparison_scope record incomplete; (3) dump still rejects a live partial read (ADR-079 leaves that choice open), so only an externally produced snapshot carries partial today; (4) a flat snapshot_to_dict document has no section header, so it carries no fact and reads unrecorded.

Closed (2026-10, Lane C stage 5). workflows.consumer_scope.read_consumer_facts now treats a FAILED consumer-import fact exactly like an unrecognised format: a REQUIRED consumer raises ConsumerUnreadableError, an ADVISORY one yields unreadable=True. Regression tests: tests/unit/workflows/test_failed_consumer_read.py. The library side followed in Lane C stage 6: parse_{elf,pe,macho}_metadata still return what they read, but each site that loses export-table facts records it (extract/parse_failures.py), and read_library_export_facts reads such a library -- or a stored block that records no parse (model.export_index.platform_block_parsed) -- as FAILED. The scoping workflow then raises LibraryExportsUnreadableError (compare --used-by: exit 1, Error: --used-by library: ...). Regression tests: tests/unit/extract/test_library_export_read_failures.py. Still open: a stored snapshot keeps no record of a partial read (a skipped .dynsym with a parsed header) -- persisting it needs a snapshot-schema field and its ADR; dump rejected every such binary tried (ELF and whole-file PE/Mach-O failures), so in practice only a snapshot produced elsewhere can carry one.