eliziff is making Mike's front end earn its keep

Un-export 23 module-internal symbols, delete 5 fully dead ones
(including the fast-diff dependency, imported solely to be re-exported
'for future work'), and collapse five local sha256 definitions, three
XML entity decoders, and two whitespace normalizers into lib/hash.ts
and lib/text.ts. officeText's decoder decoded & first, which
double-decoded entities; the shared decoder fixes that.

tsc clean; 639 backend tests pass.

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

The MCP SDK cost ~270ms of require time on every server boot even with
zero connectors configured. Load it on first connector use instead:
type-only imports stay, the Client/transport pair loads inside
withMcpClient, and the OAuth 'auth' entry point loads behind a tiny
wrapper at its two call sites. Eager module graph drops 674 -> 472
modules; boot 764ms -> 581ms.

tsconfig gains 'incremental': full backend build 8.3s cold -> 4.4s warm.

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

undici (95 modules, ~100ms) was in the eager graph solely for the SSRF
guard's pinned-DNS Agent, which is per-request work. With this and the
lazy MCP SDK the server's eager module graph is 674 -> 377 modules.

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

Every adapter duplicated the same two rituals: a ~10-line apiKey()
override-or-env-or-throw copy, and paired logRawLlmStream(...) +
rawStreamRecorder?.record(...) calls at each stream-event site.
createLlmTrace() now folds console echo and file capture behind one
always-present handle, and requireApiKey() owns key resolution.
Also drops OpenAI's redundant sse_chunk capture (the parsed sse_event
records the same bytes, matching the DeepSeek adapter's granularity).

Verified: tsc clean; 47 llm+chat tests green incl. new smoke test
proving file capture JSON and console echo end-to-end.

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

xlsx (~1 MB, installed from a pinned CDN tarball) sat in the eager
boot graph for every process that never touches a spreadsheet.
spreadsheetToLLMText is now async with a memoized dynamic import;
verified via require-cache probe that xlsx no longer loads at boot.

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

CLI default is high on effort-capable models and this machine's
settings.json pins effortLevel high, so pre-pin cc-harness rows and
judge calls were already tier-parity with claude-sonnet-4-6-high.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018bhP89cYSdEEjHNPnxacns

Citation detection/resolution goes through citationKey.ts,
caselawCitator.ts (citationAliasKeys), the corpus lookup indexes, and
the US lane (courtlistenerLocalBulk + eyecite); genuinely new shapes
expand the shared module with tests. CITATION_IN_QUERY in
a2ajPassageSearch.ts recorded as pre-existing debt. Also carries a
concurrent session's journal-footnote-pairs data note.

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

The plan still read as though a Stage 19 burn were queued and Tier C were the
last gate before it. That framing predates the four audits.

Current prior stated up front: LegalBench-RAG is NOT the main bed - no case law,
no citations, no courts, so the citation web, court hierarchy and attested
characterization are structurally unmeasurable there, and F1 showed its
retrieval numbers are substantially a query-format artifact. It gets closed out,
not extended.

Phases replaced by an explicit priority order: harden the deterministic
structure layer, then premise #1's checker-family crossing (both in flight),
then free re-tests on banked data, then the case-law bed where the thesis
actually lives, then the arms never run, then the structure round's composed
ablations. LegalBench closeout drops to last and is explicitly "not a gate for
anything": Tier C answers whether G+ctx was ever separable from plain G, and the
hold-out burn is a separate decision rather than something queued behind it.
File the CRLF defect upstream regardless.

Principle 8 (nothing retired on a deterministic proxy) now carries the full
retirement standard. Structure-round doc re-pointed at priority 6.

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