chore: release v0.5.124

[gonewton-deploy]

Automated release for version 0.5.124

This PR was created by the auto-release workflow. Merging it brings
the bumped version back into main.

What triggered this release:
Commit: c8851c8
Message: fix: tranche 4 audit remediation — smells & hygiene (spec 074) (#497)

Implements Tranche 4 ("smells & hygiene") of the audit remediation spec
(074, local), following tranches 1 (#491), 2 (#493), and 3 (#495). Ten
work items plus a code-review pass with fixes.

Changes

• Repo hygiene (H1-H6): untracked 220 accidental UUID-named
  workflow-run artifacts under crates/cli/.newton/, fixed .gitignore's
  .newton/* root-only anchoring (**/.newton/* + a
  !**/.newton/scripts/** exception), deleted three empty/abandoned
  directory skeletons and two throwaway files, deleted-or-implemented
  scripts/linter.sh, fixed architecture.md's operator table and
  docs/realtime.md's stale event-type documentation.
• Dead clap layer + operator module hygiene (S6/S13): stripped
  unused clap derives/attributes from crates/cli/src/cli/args.rs (kept
  as plain DTOs — real parsing is cli-framework's ArgSpec layer),
  removed the now-unused clap dependency, moved assessment.rs to
  workflow/grading/ (it's a helper library, not an operator).
  gh_authorization.rs was not deleted — investigation showed it's
  load-bearing (GhOperator's real authorization flow, 65 passing tests)
  despite the audit's premise that it was orphaned.
• Small fixes (P7/P13/P14): a new lint rule warns when
  model_stylesheet.context_fidelity is set to a non-default value that
  has no effect yet; purged the removed monitor CLI command from skill
  metadata; wired the previously-discarded _terminal_stop signal into a
  new WorkflowExecution.terminal_stop field (kept the lossless
  WorkflowExecutionStatus mapping from S11 intact — added a field rather
  than a new status variant).
• Unified event-sink trait (S7): DbSink/ServerNotifier were
  already consolidated to one WorkflowSink trait with two impls by an
  earlier tranche (premise correction — the audit's "collapse to one
  trait" text no longer matched current code); this item's remaining scope
  was replacing let _ =-swallowed channel-send failures with
  tracing::debug!-logged failures.
• spawn_blocking for workflow-file I/O (S10): all five
  workflow-files axum handlers (list/get/put/delete/validate) wrap their
  blocking std::fs calls and run_diagnostics' parse/validate/lint in
  tokio::task::spawn_blocking, following the codebase's existing
  outer-JoinError/inner-Result idiom.
• Typed Finding lifecycle enums (S3): FindingStatus (8 canonical
  CONTEXT.md variants), Severity, and Origin replace String fields
  on FindingItem/CreateFindingBody/PatchFindingBody.
  finding_row_to_item (the store boundary) parses DB strings via
  FromStr, treating an unparseable value as store corruption rather than
  silently coercing it. Deleted the now-redundant FINDING_STATUSES
  allowlist (it only covered 5 of 8 canonical statuses) — invalid values
  are now rejected at the type level. reconcile.rs's is_open_status
  takes &FindingStatus with an exhaustive match, the actual mechanism
  behind "the compiler owns what is_open_status approximates."
• Route colocation + pagination (S12):
  Product/Component/Repo/ModuleDependency's read routes
  (previously split across dashboard.rs/portfolio.rs from their CRUD
  in catalog.rs) are now colocated under one "catalog" OpenAPI tag.
  Added limit/offset pagination (default 100, hard cap 1000) to the
  previously-unbounded
  kpis/modules/plans/products/components/regressions list
  endpoints named by the audit.
• Capture-truncation handling + hot-path panics (S15): every
  silently-swallowed stdout/stderr capture write (let _ = write_all(...)) now tracks its failure reason and appends a [capture truncated: <reason>] marker to the artifact file — covering both
  genuine I/O failures and the pre-existing byte-cap cutoff, neither of
  which used to leave any trace. The truncation reason also surfaces on
  the task's own output JSON
  (stdout_capture_warning/stderr_capture_warning) since no per-task
  "warnings" channel exists on ExecutionContext. Replaced the two
  .expect("stdout/stderr must be piped") panics on the subprocess hot
  path with typed AppErrors.
• Rhai env() opt-in per workflow (S8, part 1):
  ExpressionEngine::default() no longer registers env() — expression
  evaluation is deterministic unless a workflow explicitly sets the new
  settings.allow_env_fn: true. Threaded through all real-execution call
  sites (child_runner.rs, newton run's loader.rs path, newton optimize's workflow driver) while validation/lint/explain paths stay
  deterministic by design.
• LlmAdjudicator/AgentClient trait injection (S8, part 2):
  ReconcileOperator, GraderAgentOperator, and ChangeRequestOperator
  no longer call aikit_sdk::AgentRunner/Pipeline directly inside
  execute() — a new llm_client.rs module defines the two seams, each
  with a real implementation that moves the prior spawn_blocking body
  unchanged, and each operator gains an additive
  with_adjudicator/with_agent_client constructor (the plain ::new()
  and every existing test are untouched). This is the actual enabler for
  testing B2's failure path (adjudication failure must fail closed with
  zero Finding mutations) — previously only reachable via a real LLM call
  failing.
• Shared FK-validation service layer (S5): investigation found the
  audit's premise didn't fully hold —
  create_component_db/create_repo_db/create_module_db/create_eval_run_db/create_grade_db
  already validate their FKs transactionally and return a clean 404, so
  the axum handlers and dispatch_data's real write path were never
  actually duplicating broken validation. The one real, still-live
  duplication was narrower: the CLI's --dry-run block hand-rolled its
  own separate copy of the same "does this FK exist" checks. Extracted
  those into crates/backend/src/service.rs (validate_*_fks, five
  resources) and wired --dry-run to call them, collapsing ~115 lines of
  hand-rolled match arms. Deliberately not wired into the real write paths
  (would reintroduce a TOCTOU gap for no behavioral gain).

Review (8-angle, high effort)

8 finder angles (line-by-line, removed-behavior, cross-file tracer,
reuse, simplification, efficiency, altitude, conventions) plus targeted
verification. Fixed on-branch:

• env() opt-in didn't reach real execution (found by the
  cross-file tracer, independently confirmed by direct code reading):
  apply_default_pipeline — which runs the
  macro-expansion/include_if/template-interpolation transforms — was
  called before the settings-aware ExpressionEngine was built in
  child_runner.rs/optimize.rs, and each of those three transforms
  constructed its own always-deterministic engine. A workflow that opted
  in via settings.allow_env_fn: true would still fail to load if it used
  env() inside a template string, macro argument, or include_if
  condition (as opposed to a task's own $expr param, which worked
  correctly). Threaded allow_env_fn through apply_default_pipeline's
  signature into all real-execution call sites — this also caught a third
  genuine execution path, loader.rs's load_and_lint_workflow (used by
  newton run despite its lint-sounding name), that the original S8 work
  missed. Validation/lint/explain paths explicitly stay deterministic
  (documented at each call site). New regression test proves both
  directions: fails without the opt-in, works with it.
• Finding enum malformed-value error shape: typing
  CreateFindingBody/PatchFindingBody's status/severity/origin as
  enums moved invalid-value rejection into serde deserialization inside
  axum's Json<T> extractor — before the handler runs, and before this
  API's usual ApiError JSON envelope gets applied. A bad value on POST /findings/PATCH /findings/{id} fell through to axum's default
  plain-text rejection instead. Added AppJson<T>, a drop-in
  FromRequest wrapper mapping JsonRejection to a 422 ApiError,
  adopted by the two Finding handlers that surfaced the bug (kept as a
  reusable primitive; adopting it repo-wide is a documented follow-up, not
  needed by any other endpoint today).
• /repos and /module-dependencies missing pagination: these two
  endpoints were relocated into catalog.rs by the same S12
  route-colocation work but never got the limit/offset treatment their
  newly-adjacent siblings received. Applied the same pattern.
• S15 capture-truncation logic duplicated between command.rs and
  sdk.rs: a coverage-gap-closing commit had already extracted
  write_capture_chunk inside sdk.rs to deduplicate three copies there,
  but left command.rs's two structurally-identical inline copies
  untouched — confirmed by two independent finder angles, and by git
  history showing the exact same commit introduced both files'
  duplication. Moved the helper to artifacts.rs (which both files
  already import from) so both share one implementation.
• NEWTON-0011 doc-comment gap: the three new
  FindingStatus/Severity/Origin::as_str() methods had no /// doc
  comments.

Investigated and refuted or deliberately deferred:

• Four CLI arg-parsing gaps (--port/--format
  silent-default-instead-of-error, --parallel-limit/--timeout/--last
  negative-value integer-wraparound bypassing validation) were flagged by
  two finder angles reading adjacent code for context — all four confirmed
  real, but confirmed via git diff main...HEAD --stat to live in files
  this branch never touches (framework_setup/mod.rs,
  commands/serve.rs, commands/workflow.rs), pre-existing on main
  since tranche 3. Out of scope for this PR; worth a follow-up ticket.
• A claimed change_request_op.rs severity-fallback behavior change
  (garbage min_severity silently landing on a more permissive rank than
  before) was refuted — the pre- and post-enum fallback both land on rank
  4; the reviewer conflated it with an unrelated function's
  .unwrap_or(2).
• RealLlmAdjudicator duplicating RealAgentClient's spawn_blocking
  body instead of delegating to it, and workflow_files.rs's 5x-repeated
  spawn_blocking+JoinError triad, are real but lower-severity — left
  as documented follow-ups rather than expanding this PR further.
• The Finding/ChangeRequest FK-validation asymmetry (they still lack
  the pre-check the 5 covered resources have, so a bad FK still surfaces
  as a raw 500 rather than a clean 404) is real but pre-existing, not
  introduced by this branch — the S5 commit's premise only checked the 5
  resources it covers, not these two. Documented follow-up.
• Pagination fetching the full unbounded result set and slicing in Rust
  rather than pushing LIMIT/OFFSET into SQL is a deliberate,
  already-documented tradeoff from the S12 work (pushing to SQL would
  require changing the BackendStore trait and its CLI/grading call sites
  — out of scope for a hygiene fix on tables that aren't expected to be
  large).

Verification

• All suites green across newton-types/backend/core/cli; fmt + clippy
  -D warnings clean; full cargo test --workspace green after every
  commit, independently re-verified (not just trusting subagent
  self-reports) at each of the 10 work items plus the coverage-gap and
  review-fix commits.
• Diff coverage 94.45% (1567/1659 instrumented added lines vs
  main) — an initial pass measured 88.54%; closed the gap with targeted
  tests (the AgentClient/LlmAdjudicator real-implementation error
  paths via the established "non-runnable engine" test trick, the S15
  truncation-marker branches via a real over-cap subprocess and a
  helper-level unit test, FindingStatus/Severity/Origin round-trip
  tests, and spawn_blocking JoinError-mapping tests via a panicking
  store test double).
• 8-angle code review with targeted 1-vote verification on ambiguous
  findings; 5 confirmed issues fixed on-branch (above).
• One commit (context_fidelity lint / monitor purge / terminal_stop,
  originally 90 characters) exceeded NEWTON-0009's 72-char subject limit —
  fixed via scripted git rebase -i reword before push, content
  byte-identical.

Known follow-ups (documented, not blocking)

• Four pre-existing CLI arg-parsing validation gaps discovered
  incidentally during review (silent-default-on-invalid-value for
  --port/--format, integer-wraparound-bypasses-validation for
  --parallel-limit/--timeout/--last) — live in code this branch
  never touches, predate tranche 4 entirely; worth a dedicated follow-up.
• Finding/ChangeRequest creation still has no FK pre-check (unlike
  the 5 resources S5 covers) — a bad componentId/repoId/kpiId on
  POST /findings surfaces as a raw 500 rather than a clean 404.
• RealLlmAdjudicator::adjudicate duplicates
  RealAgentClient::run_pipeline's spawn_blocking body rather than
  delegating to it with adjudication-specific template/schema/parsing — a
  ~45-line reduction is available.
• workflow_files.rs's spawn_blocking+JoinError three-way match is
  repeated at all 5 handler call sites; only the join-error arm is truly
  shared logic across them.
• AppJson<T> (the new Json<T> rejection wrapper) is adopted only by
  the two Finding handlers that surfaced the bug; the rest of this API's
  many Json<T> extractors could adopt it for consistency, though none
  have this specific typed-enum-rejection issue today.
• Pagination's limit/offset clamp logic
  (DEFAULT_LIST_LIMIT/MAX_LIST_LIMIT/clamp helper) is duplicated
  across catalog.rs/dashboard.rs/plans.rs rather than shared from
  one location.
• service.rs's FK-existence checks fetch a full row via
  get_product/get_component/get_repo rather than the cheaper
  row_exists probe already used for the same purpose elsewhere in the
  store layer.