Statement coverage to 82% with an enforced floor; resolve the four zero-importer modules (#196)

[cdeust]

Closes #196.

What this PR does

1. Wires condense_assembled_context as pg_recall.assemble_context's missing truncation-awareness pass — the stage assembler's three phases each enforce only their own sub-budget, so their sum could exceed the caller's token_budget; the new pass is a byte-identical no-op when the sections fit and priority-condenses (own > adjacent > summaries) through decomposer.assemble_prompt when they don't. This also gives decomposer.py its production call site (partially addresses context_assembly: decomposer + active_retrieval have no call site #201).
2. Raises statement coverage over mcp_server from 74.83% to 82.02% (CI coverage job, run 30316539703, coverage.py 7.15.2) via contract tests for: stage_detector, decomposer, ppr_traversal, active_retrieval, coverage, warning, concept_emerger, wiki_view_executor, the pg_recall wiring, agent_briefing, wiki_rule_engine, wiki_file_doc_skeleton, wiki_coverage_dashboard, session_start (the issue's largest named gap, 36% → 82%), batch_sinks/pooled_sink, and viz_client. Measured locally with the same command on the SQLite backend the tree still clears the bar on its own: 80.14% (29,534 of 36,854) with every PG-only test failing-by-environment.
3. Enforces the floor: --cov-fail-under=82 in the CI coverage job.
4. Flips test_statement_coverage80 to Met in .bestpractices.json and updates docs/ASSURANCE-CASE.md § "What this assurance case does NOT claim".
5. Isolates domain_mapping's dev-root scan in conftest — the transitive os.walk over real repo trees killed local --cov runs at the 300 s per-test timeout; CI isolation was accidental (no dev roots on runners), now it is designed.

The four zero-importer modules (criterion 1)

Module	Resolution	Evidence
infrastructure/git_diff.py	Deleted on main by #204 — behaviour proven to live in cortex-viz's independent git_diff_engine.py; its one shared consumer (shared/subprocess_safe.py) docstring names the removal	git log --follow shows removal in 1572f61; no importer existed
shared/memory_types.py	Deleted on main by #204 — scaffolding types with no call site ever built	removal in 1572f61
core/context_assembly/condensers.py	Caller built: #204 wired condense_memory_content into stage_phases; THIS PR completes the wiring with condense_assembled_context inside pg_recall.assemble_context	tests_py/core/test_pg_recall_assemble_context_condensation.py (3 tests), tests_py/core/context_assembly/test_condensers.py (33 tests)
hooks/preemptive_context.py	Caller exists outside the import graph: registered as a PostToolUse hook in .claude-plugin/plugin.json:79; covered by tests_py/hooks/test_preemptive_context.py (22 tests, merged in #204)	zero importers is the expected shape for a hook entry point

hooks/agent_briefing.py (SubagentStart) and hooks/session_start.py (SessionStart) share the same entry-point shape — registered in plugin.json, zero importers by design — and are now covered by 29 and 64 tests respectively.

Mutation triage on the changed function (§12)

Scoped mutmut run (only_mutate=condensers.py, 352 mutants) after adding the new tests: 0 non-equivalent survivors in condense_assembled_context — 13 killed by three added contract tests (exact-budget byte-identity; every-section-yields with banner-key labels; adjacent-yields-to-own priority), 1 documented equivalent (summaries priority 3→4: priorities are pure sort keys and the mutation preserves the strict order own < adjacent < summaries — behaviour identical). The 125 survivors in the PRE-existing condenser functions (code this PR does not change) are the sweep-tier backlog, filed as #228. The changed lines in pg_recall.py are pinned by the three wiring tests (over-budget condenses; token_budget=None never invokes the pass; under-budget output unchanged).

Completion Ledger

§13.1 checklist

Item	State	Evidence
A1 happy paths	done	full suite: 6029 passed / 3 skipped (CI run 30316539703); e.g. test_over_budget_result_is_condensed, test_assembled_context_under_budget_matches_stage_assembler_format
A2 edge cases	done	empty/whitespace sections (test_assembled_context_all_empty_returns_empty_string, test_assembled_context_whitespace_only_input_treated_as_empty), boundary (test_assembled_context_exactly_at_budget_is_byte_identical), oversize (test_assembled_context_over_budget_condenses_every_section), unicode-free blobs + sentence content both exercised
A3 failure paths + signals	done	every degraded arm asserts its emission: test_failed_agent_scoped_query_is_logged_and_team_pass_still_runs, test_failed_team_query_is_logged_and_returns_partial_results, test_unreachable_postgres_is_skipped_with_a_signal, test_unreadable_page_is_skipped_not_fatal, test_unwritable_dashboard_is_skipped_and_the_batch_continues, test_unwritable_index_does_not_lose_the_dashboards, test_write_dashboards_registry_failure_writes_nothing, test_invalid_regex_silently_does_not_match
A4 trust boundaries	done	hook stdin gates: test_main_ignores_malformed_json, test_main_ignores_empty_stdin, test_main_ignores_an_interactive_terminal; non-string tags: test_tag_match_is_case_insensitive
A5 invariants stated	done	condense_assembled_context docstring states pre/postconditions incl. the banner-length caveat; unreachable-arm proof comment at the condense_assistant_message final return
A6 idempotency	N/A	pure functions (condensers, rule engine, skeleton) and read-only queries; write_dashboards re-run overwrites deterministically (test_write_dashboards_emits_one_page_per_domain_plus_index)
B1-B3 concurrency	N/A	no concurrency touched; all new code is synchronous pure logic + per-call DB reads
C1 scalability	done	dashboard caps its uncovered listing at 30 (test_resolved_source_root_reports_file_coverage_and_caps_the_list); keyword extraction caps at 8; briefing capped at _MAX_MEMORIES
C2 resource lifecycle	done	connection closed on every arm: conn.close asserted in test_briefing_prints_header_marker_and_one_line_per_memory and test_no_relevant_memories_is_skipped_and_connection_closed; temp dirs via tmp_path
C3 hot paths	N/A	no hot path touched; condensation pass is a no-op (total <= token_budget fast path) in the common case, asserted byte-identical
D1-D3 security	done/N/A	briefing SQL uses parametrized queries only (asserted through call_args); no secrets; no new privileges; no query built from user strings
E1 API compat	done	additive only — assemble_context return keys unchanged; raw phase fields stay uncondensed (test_over_budget_result_is_condensed asserts this consumer contract)
E2 downstream consumers	done	named: benchmark evaluators matching selected_memories against raw phase fields — pinned by the same test
E3 persisted data	N/A	no format change
E4 cross-platform	done	no path/process/encoding logic added beyond tmp_path-based tests; suite green on Windows leg (CI run 30316539703)
F1 failure signals asserted	done	see A3 row
F2 degraded modes explicit	done	marker-less briefing on failed receipt: test_failed_receipt_degrades_to_marker_less_briefing; "(no source root resolved)" line: test_fresh_domain_renders_all_slots_as_queued
G1 path ledger	done	table below
G2 regression tests	N/A	no bug fixed in this PR (coverage + wiring work)
G3 deterministic tests	done	tmp_path everywhere, no sleeps, no shared state; monkeypatched Path.home; suite passes under default parallel CI matrix
G4 negative assertions	done	absence checks throughout: connect.assert_not_called() in the briefing gates, mock_condense.assert_not_called() in test_none_token_budget_skips_condensation, assert "Related Prior Context" not in result, receipt-marker absence in test_failed_receipt_degrades_to_marker_less_briefing
G5 suite + gates	done	local: 5895 passed / 176 skipped, plus 17 failed + 27 errors that are pre-existing PG-unreachable artifacts (identical on unmodified main via paired control; tracked by #220); CI: 6029 passed / 3 skipped (run 30316539703); ruff format/check clean on touched files; pyright ratchet PASS (no blocking rule regressed)
H1 standards	done	no layer violation (tests + one core-to-core wiring); sizes within caps; sourced constants (banner threshold cites decomposer defaults)
H2 readable/simplified	done	no dead code added; unreachable arm documented with proof rather than deleted (future-fallback rationale at use site)
H3 conventions	done	function-style tests matching each target file's neighbours; one language per file
H4 CHANGELOG/docs	done	advertised test count synced to 6115 across README/CLAUDE/CONTRIBUTING/ASSURANCE-CASE/bestpractices; assurance-case §6 updated
H5 commit hygiene	done	conventional commits; formatting isolated in its own commit
H6 CI green	done	run 30316539703 on the pushed tree
H7 boy-scout (§14)	done	coverage.xml gitignored (seen untracked artifact); pre-existing SQLite-mode failures reproduced on unmodified main (paired control, 16/16 identical) — in scope of open issue #220; pre-existing mutation survivors filed as #228; no other warnings surfaced

Code-path ledger (new paths introduced by this diff)

Path	Asserting test
condense_assembled_context fast path (fits)	test_assembled_context_under_budget_matches_stage_assembler_format, test_assembled_context_exactly_at_budget_is_byte_identical
section omitted when empty/whitespace	test_assembled_context_omits_empty_sections, test_assembled_context_whitespace_only_input_treated_as_empty
all-empty → ""	test_assembled_context_all_empty_returns_empty_string
over-budget condensation path	test_assembled_context_over_budget_result_fits_within_budget, test_assembled_context_condensation_engages_and_reduces_content
priority ordering own>adjacent>summaries	test_assembled_context_own_stage_survives_more_than_summaries, test_assembled_context_adjacent_yields_budget_to_own_stage
truncation banner emission + per-key labels	test_assembled_context_emits_truncation_warning_banner, test_assembled_context_over_budget_condenses_every_section
over-budget template excludes absent sections	test_assembled_context_over_budget_excludes_missing_sections
assemble_context wiring: budget set + over	test_over_budget_result_is_condensed
wiring: budget set + fits (no-op)	test_under_budget_result_unchanged
wiring: token_budget=None skip	test_none_token_budget_skips_condensation
condensers unreachable no-prose-budget return	# pragma: no cover with written single-segment proof at the use site (provably unreachable; kept as fallback)

New-test files for previously-uncovered modules map one test per behaviour; per-file counts: session_start 64, agent_briefing 29, wiki_view_executor 57, concept_emerger 51, stage_detector 30, ppr_traversal 27, decomposer 23, active_retrieval 18, wiki_file_doc_skeleton 17, coverage 13, wiki_coverage_dashboard 16, wiki_rule_engine 15, viz_client 12, batch_sinks 7, warning 5, pg_recall wiring 3, condensers +13.

Deferrals (each with its issue number)

• Pre-existing mutation survivors in the untouched condenser functions → condensers.py: 125 pre-existing mutation survivors outside the #196-changed function #228 (outside this change's blast radius: no behaviour of those functions changed here).
• Pre-existing full-suite-under-SQLite failures (16 tests, reproduced identically on unmodified main via paired control) → already tracked as SQLite backend full-parity: CI runs 1 file, 35 tests fail on an isolated SQLite store #220.

🤖 Generated with Claude Code