[e54-bot]

What changed

wright_driver::connect::connect_provider is now the single LPP connect path every Wright provider client uses. It owns:

• provider lookup — the session registry plus the opy id's configured resolve(); the LSP installed() fallback is removed per the owner decision (every client resolves);
• version negotiation with a floor (LPP §19.2) — initialize at the caller's preferred version; on protocol-version-mismatch, shut down, respawn, and retry once at the highest mutually supported version at or above the ConnectFloor (entry targets ≥ 1.1, directory targets ≥ 1.3 per LPP client: capability subsets rejected, repeated provider diagnostics dropped, UTF-16 columns, client name #638);
• failure classification — all failures surface as the typed ProviderError so callers map them through the shared provider-refusal-<code> classification.

Callers migrated to the shared path:

• CompilerSession::load_from_source_provider (the driver's spawn + 1.4-preferred initialize + floored retry, now via connect_provider);
• CompilerSession::run_provider_flow — fix context (provider_fix_validation) and ProviderRename/language-service rename — which previously called plain initialize with no negotiation;
• opy_lookup (preferred 1.5, ConnectFloor::Entry);
• convert_opy (preferred 1.0);
• LanguageService's OPY provider (connect_opy_provider deleted; the service now reuses its lazy CompilerSession's connect_provider).

The LSP's parallel connect_opy_provider/mutual_protocol_version copy is deleted, and the language service's refusal mapping now uses the driver's provider_failure_code — a refusal reports provider-refusal-<code> on both the CLI and language-service surfaces instead of the raw refusal code.

CompilerSession::language_provider remains as the spawn-only seam for hosts that initialize the provider themselves (its doc points at connect_provider), and ToolService::language_provider keeps delegating to it.

Verified

• cargo fmt --all -- --check — clean
• cargo clippy --workspace --all-targets --all-features -- -D warnings — clean
• cargo test --workspace --all-targets --all-features — all binaries green (e.g. wright-cli cli.rs 124/124, wright-driver 22 test binaries, wright-language 14+6, wright-lsp 25/25)
• cargo check -p wright-wasm --target wasm32-unknown-unknown — clean (the connect path stays native-only per ADR-0037)
• New language-service test (an_opy_provider_negotiates_and_classifies_refusals_like_the_driver) spawns a provider that rejects the preferred protocol version, verifies the §19.2 retry at the highest mutual version, and asserts an lpp/check refusal surfaces as provider-refusal-check.tooLarge — matching the driver-side retry test (a_version_mismatch_retries_at_the_highest_mutual_version).
• LSP/language-service boundary tests pinned offline (WRIGHT_OPY_PROVIDER_BASE_URL unreachable / nonexistent explicit executable) so the resolve-only contract is deterministic.

Not verified

• No real opy-rs provider binary was exercised end-to-end on this machine; the negotiation is covered by the Python fake-provider tests on both paths.
• ui/ gates (npm ci && npm run check && npm test && npm run build && npm run check:size) — there is no ui/ directory in this repository, so those checks do not apply.

Fixes #1004

[e54-bot]

All three r1 findings addressed in e2d731be (fix: stop classifying initialize refusals and session failures as provider refusals):

1. lookup.rs initialize refusal: the connect Err arm no longer reclassifies an lpp/initialize LPP refusal as an unavailable capability gap. Refusals now carry the shared provider-refusal-<code> classification; only resolution, spawn, process, and version-negotiation failures stay unavailable. Verified live: a provider answering lpp/initialize with refusalCode: init.tooBusy now returns {"error":{"code":"provider-refusal-init.tooBusy","message":"LPP error: refusal: refusal: busy"}} instead of the Ok/unavailable envelope. Regression test: opy_lookup_carries_an_initialize_refusal_with_its_code.

2. service.rs session-construction classification: CompilerSession::new diagnostics are no longer fabricated into ProviderError::lpp(Refusal). OpyProviderState::Unavailable now latches the classified (code, message), so a construction failure surfaces its own code (invalid-selection) on .opy diagnostics and rename instead of provider-refusal-invalid-selection. Regression test: a_session_construction_failure_keeps_its_own_code.

3. language-services.md: the OPY provider session docs now describe the resolve() path — explicit executable, installed store, or bootstrap download — matching the driver, instead of claiming "never a download".

Gates: cargo fmt --all -- --check, cargo clippy -p wright-driver -p wright-language -p wright-lsp --all-targets --all-features -- -D warnings, cargo test -p wright-driver --test lookup -p wright-language --test service (all green), git diff --check — all pass. service.rs stays at 842 lines (no size-guard growth).

Ready for follow-up review.

[e54-bot]

Review findings addressed in e2d731be, pushed to gh-1004-one-lpp-connect.

• Finding 1 (fixed) — crates/wright-driver/src/lookup.rs: the Err arm of connect_provider no longer reclassifies an lpp/initialize LPP refusal into an unavailable capability gap. Resolution, spawn, process, and version-negotiation failures (NotConfigured, Spawn, Io, Exited, Local, ProtocolVersionMismatch, plus Lpp answers of kind protocolVersionMismatch/capabilityUnavailable) stay unavailable; every other ProviderError — including a Refusal — now goes through carried_refusal(provider_failure_code(&error), …), so init.tooBusy publishes provider-refusal-init.tooBusy. Regression test opy_lookup_carries_an_initialize_refusal_with_its_code added; live CLI check: wright lookup --opy-provider <refusing script> --language opy --format json hud → {"error":{"code":"provider-refusal-init.tooBusy",…}}.
• Finding 2 (fixed) — crates/wright-language/src/service.rs: CompilerSession::new failures are no longer fabricated into an LPP Refusal. The latched connect failure is a classified (code, message) pair — provider errors classified once through provider_failure_code, session-construction diagnostics carried verbatim — so .opy diagnostics and rename report invalid-selection, not provider-refusal-invalid-selection. Regression test a_session_construction_failure_keeps_its_own_code added.
• Finding 3 (fixed) — docs/language-services.md: the OPY provider session paragraph now says it resolves through resolve() — explicit executable, installed store, or bootstrap download — matching the driver, instead of claiming "never a download".

Verification: cargo fmt --all -- --check, cargo clippy --workspace --all-targets --all-features -- -D warnings, cargo test --workspace --all-targets --all-features (all pass), git diff --check, plus the live CLI falsification above.

The PR is ready for follow-up review.

[Teakowa]

LGTM after independent review.

[e54-bot]

Merge conflicts with main resolved in dc842bc0 (merge: resolve conflicts with main (typed findings, store helpers)), pushed to gh-1004-one-lpp-connect. The PR is now MERGEABLE.

Resolution, keeping both intents:

• crates/wright-language/src/service.rs — one conflicted file, two hunks.
  • Imports: merged — kept Severity + DiagnosticSeverity (main's typed-findings refactor: keep findings typed through driver results #1015) alongside this branch's connect path; dropped ResolvedOpyProvider and ProviderError (main's free connect_opy_provider that used them was superseded by this branch's shared-connect method).
  • Dead helpers: both lpp_severity (this branch's, made obsolete by typed severity) and main's local provider_failure_code had no remaining callers after the union — both removed; the diagnostics boundary uses the shared wright_driver::source_provider::provider_failure_code.
• Kept this branch's intent: the language service still connects through CompilerSession::connect_provider — no installed() fallback, provider-refusal-<code> refusal classification, session() surfaces construction diagnostics verbatim.
• crates/wright-driver/src/lib.rs auto-merged: pub mod connect / ConnectFloor / ConnectedProvider coexist with main's pub mod store promotion.

Verification after the merge (all green):

• cargo fmt --all -- --check
• cargo clippy -p wright-driver -p wright-language -p wright-lsp -p wright-cli --all-targets --all-features -- -D warnings
• cargo test -p wright-driver -p wright-language -p wright-lsp --all-targets — 0 failures
• cargo test -p wright-driver --test lookup -p wright-language --test service — all pass, both regression tests included
• git diff --check
• Live CLI falsification: wright lookup --opy-provider <refusing script> --language opy --format json hud still returns {"error":{"code":"provider-refusal-init.tooBusy",...}}

Ready for a conflict-resolution re-review.

[Teakowa]

LGTM after independent review.