Fix report_error! demotions of typed errors into extra: (APP-5522)

[warp-agent-staging]

Description

Fixes APP-5522: report_error! call sites that demoted a real, typed error into extra: { "error" => ... } (or stringified it via anyhow!("{e}")/interpolated it into the grouping message), contradicting .agents/skills/logging-and-error-reporting/SKILL.md ("Choosing the form" rule 1 and the Anti-patterns block). Reported by Andrew Carlson via Slack: the flagged line.

The demotions were introduced by #13483 (the mass log::error! → report_error! migration), which also added the skill that forbids them. A sibling task is separately hardening the review-prompting that let this class of regression through; this PR is the code cleanup only — it does not touch anything under .agents/skills/.

Note on the originally-flagged line

An unrelated concurrent PR (#15287, merged just before this branch was cut) already fixed the two spawn_tui_driver draw-error sites in crates/warpui_core/src/runtime/mod.rs (the one the complaint linked, plus the invalidation-callback twin), wrapping them in .context() and adding ReportErrorLogMode::OncePerRun. This PR leaves those two alone (verified the landed form is correct) and instead fixes the third, still-broken site in the same file — the terminal-event reader — which #15287 didn't touch.

Categories fixed

• Category A (typed Error stuffed into extra: only): crates/warpui_core/src/runtime/mod.rs (event reader), crates/http_server/src/lib.rs, crates/warpui_core/src/core/app.rs, crates/warpui_core/src/integration/mod.rs, crates/warpui_extras/src/user_preferences/registry_backed.rs, app/src/ai/geap_credentials.rs.
• Category B (String/non-Error extra: family): upgraded interim anyhow!("{e}") wraps where a Result was in hand (commit.rs, default_terminal/mod.rs), removed a double-report in the Docker sandbox environment-prep path (the inner path already reports the typed error — the outer sink now just logs), and switched terminal_pane.rs's remote-child launch failure to report the typed error with conversation_id kept as incidental extra:. Also upgraded the two notification-permission "unknown error" sites (terminal/view.rs, workspace/view.rs x2) to treat the message as the payload instead of extra:.
• Category C (message interpolation): crates/warp_tui/src/terminal_session_view.rs — stopped putting {error} in the grouped message for the two settings-persistence paths.
• Category D (anyhow!("{e}") stringify sites, defeats is_actionable): rewrote to preserve the typed chain via anyhow::Error::new(e).context(..) (or e.context(..) when already anyhow, or &e when the value is still needed afterward and can't be cloned). Two Display-only, unregistered error types (LoadGeapCredentialsError, PrepareRemoteChildLaunchError) and one unrelated Display-only type discovered along the way (HexColorError) were upgraded to real thiserror::Error types so they can be reported typed instead of stringified — see the review-fix note below on keeping LoadGeapCredentialsError's Display static.

Left alone (verified, not violations)

• local_tty/terminal_manager.rs:519 — JoinHandle::join()'s error is Box<dyn Any>, not std::error::Error.
• app/src/pane_group/child_agent/mod.rs:222 — error_message is a request field (String), not a caught Error; static message + incidental extra: is correct under rule 4.
• fetch_conversation.rs:99, code_diff_view.rs:1048 — the wrapped functions (materialize_tasks_to_yaml, restore_diff_base) already return Result<_, String>; there's no typed chain to preserve, so the existing anyhow!("{e}") form is already correct.
• crates/warpui/src/windowing/winit/event_loop/mod.rs:1952 — SoftKeyboardManager::new's error is JsValue, not std::error::Error (and not Send/Sync).
• The five Category E lookalikes named in the issue (image.rs:292, cache_setup.rs:144, response_stream.rs:693/:708, experiments/mod.rs:302/:321) were re-checked and are legitimately incidental data, not demotions.

Review fixes (round 2)

An adversarial review of the first version of this PR found that a form-correct rewrite can still ship user content to Sentry if the now-preserved Display/Debug itself carries it — the no-secrets/no-PII rule outranks form correctness. Four findings, all now fixed on this branch:

1. LoadGeapCredentialsError's derived Display interpolated detail, which carries the raw Google STS/IAM response body. Its messages are now static; detail/status stay on the type for user_facing() (which already never echoed them), but Display no longer does either. Updated the now-stale test comment that asserted no Display existed.
2. file_outline/native.rs:86 — I'd classified this as a harmless non-Error (true), but the retained anyhow!("{e:?}") was still uploading the undelivered outline map's Debug, which includes source symbol names and comments. Now a static message, throttled with OncePerRun since it can recur once per outline build.
3. The deferred wasm SoftKeyboardInput send-failure site (event_loop/mod.rs:1940) is now fixed rather than deferred: EventLoopClosed<CustomEvent> does implement std::error::Error, and CustomEvent is Send + Sync, so it's reported typed (anyhow::Error::new(e).context(..), static Display) instead of via {e:?}, which could otherwise upload the user's typed keyboard text.
4. response_stream.rs's conversion-failure path reported the same error twice — once directly, then again via report_request_failure right after. Removed the standalone report; report_request_failure is now the single sink.

Linked Issue

• APP-5522 (this is a factory:wilson-labeled hygiene issue, not ready-to-spec/ready-to-implement)

Testing

No behavior change for users — this only changes what gets reported to Sentry and how it's classified/grouped/redacted. Verified with:

• ./script/format --check — clean.

• cargo clippy --all-targets --tests -- -D warnings scoped to every touched package and its relevant feature combinations (warp, warp --features local_tty, ai, warp_core, http_server, warpui_extras, warpui_core --features tui,integration_tests, warp_tui, warpui) — clean.

• cargo check on all of the above — clean.

• cargo nextest run on all of the above — all passing except a handful of pre-existing, environment-dependent failures unrelated to this change (missing bundled SVG assets, a sandboxed nsc command being denied, a WARP_API_KEY env var leaking into a CLI-parsing test, and font-layout bidirectional-text assertions); none touch files this PR modifies.

• Two Windows-only files (registry_backed.rs, single_instance_manager.rs) could not be compiled in this Linux sandbox (no Windows cross toolchain, and no permission to add a Rust target); reviewed by hand against each dependency's documented trait impls instead. The wasm-only event_loop.rs fix (finding 3 above) is likewise unverified by a wasm build for the same reason — implemented per the review's confirmed analysis of CustomEvent's Send + Sync bound.

• Did not run the full ./script/presubmit (workspace-wide clippy/tests) due to sandbox time/disk constraints; ran the equivalent checks scoped to every touched crate instead.

• I have manually tested my changes locally with ./script/run (not applicable — no UI change; verified via targeted cargo check/clippy/nextest above)

Agent Mode

• Warp Agent Mode - This PR was created via Warp's AI Agent Mode

CHANGELOG-NONE

[warp-agent-staging]

This PR was generated with Warp.

Comment @warp-factory on this PR to send it follow-up work.