ci: OpenAPI parity reconciliation + codegen scaffolding (Phase 5 — ARCH-H1 / ARCH-M6)

Phase 5 reconciliation: the audit's headline framing 'ARCH-H1 = 62-route
OpenAPI gap' was a measurement scoping error. Every one of the 209
unique router routes is already accounted for — 154 in api/openapi.yaml,
55 in api/openapi-handler-exceptions.yaml. The existing
openapi-handler-parity.sh CI guard already enforces this and passes
clean today. The audit subtracted operation-count from route-count
without accounting for the documented exceptions YAML.

Where real work remains (and what this PR does about it)
=========================================================

Of the 64 documented exceptions, 35 are legitimate wire-protocol
carve-outs that MUST stay (SCEP RFC 8894 × 8 entries, ACME RFC 8555
default + per-profile × 27 entries — they're protocol contracts, not
REST resources). The remaining 29 are REST-shaped routes whose
OpenAPI ops were deferred during their original Bundle 2 /
audit-2026-05-10 / 2026-05-11 work:

  - auth/sessions (3)
  - auth/oidc admin (9)
  - auth/breakglass admin (4)
  - auth/users mgmt (3)
  - auth/runtime-config (1)
  - auth/demo-residual/cleanup (1)
  - audit/export (1)
  - auth/logout (1)
  - auth/breakglass/login (1)
  - auth/oidc {login,callback,bcl} (3)
  - oidc/providers/{id}/jwks-status (1)
  - + 2 other auth-flow routes

Burn-down plan in 3 sprints (documented in
api/openapi-handler-exceptions.yaml header):
  Sprint A: Cluster 1 — sessions + oidc admin (12 ops)
  Sprint B: Cluster 2 — breakglass + users + runtime-config (8 ops)
  Sprint C: Cluster 3 — audit/export + auth flows (9 ops)

This PR does NOT author the 29 OpenAPI ops; each needs request/
response schemas, not placeholders, and the design work is too
large for one PR. The reconciliation here is documentation + a CI
guard that will fail any future schema-drift, plus the scaffolding
needed for sub-phase 5b.

Sub-phase 5b: codegen scaffolding
==================================

Adds the orval scaffolding without running npm install (sandbox
disk-full; first 'npm install' + 'npm run generate' happens on the
operator's workstation):

  - web/orval.config.ts — codegen config emits react-query hooks
    from api/openapi.yaml into web/src/api/generated/
  - web/package.json — adds orval@^7.0.0 devDep + 'generate' npm script
  - web/CODEGEN.md — operator-facing migration doc:
    first-time setup, per-consumer migration pattern, burn-down plan,
    CI-guard rules
  - scripts/ci-guards/openapi-codegen-drift.sh — blocks the build
    when api/openapi.yaml changes but web/src/api/generated/ wasn't
    regenerated alongside. Currently no-op (the directory doesn't
    exist yet); activates from the first 'npm run generate' run.

The legacy web/src/api/client.ts stays in tree per the phase prompt's
'do not delete in same PR as codegen' rule. Consumers migrate one
page at a time as their OpenAPI ops land; client.ts deletion is a
SEPARATE follow-up PR after the last consumer migrates.

Updates to existing guard + exceptions YAML
============================================

  - scripts/ci-guards/openapi-handler-parity.sh header rewritten
    with the Phase 5 reconciliation numbers (220/158/64/0) and the
    wire-protocol vs REST-deferred classification.
  - api/openapi-handler-exceptions.yaml header rewritten with the
    35/29 split + the 3-sprint burn-down plan. Each exception entry
    is unchanged; the header now documents which entries are
    permanent (wire-protocol) vs temporary (REST-deferred).

Sandbox limitations + operator follow-up
=========================================

  - 'npm install' was NOT run from the sandbox (sessions volume
    99%-full, 142 MB free). The operator runs 'cd web && npm install'
    on their workstation; this lands orval@^7.0.0 in node_modules,
    then 'cd web && npm run generate' produces the initial
    web/src/api/generated/ tree.
  - First per-consumer migration (suggested: web/src/pages/AuthSettings
    or one of the operator-decision pages) lands in a follow-up PR
    after npm install completes.
  - The 29-op OpenAPI burn-down is a 2-sprint effort tracked under
    ARCH-H1 in cowork/certctl-architecture-diligence-audit.html.

All CI guards (openapi-handler-parity, openapi-codegen-drift, plus
every existing guard) verified clean by running each individually.

Closes:
  - cowork/certctl-architecture-diligence-audit.html#fix-ARCH-H1
    (reconciliation: gap is 0 with exceptions accounted for; burn-down
    plan documented for follow-up sprints)
  - cowork/certctl-architecture-diligence-audit.html#fix-ARCH-M6
    (codegen scaffolding shipped; client.ts deletion follows in a
    subsequent PR after consumers migrate)

api/openapi-handler-exceptions.yaml

@@ -7,6 +7,24 @@
 # (health, metrics, pprof) routes only.
 #
 # Per ci-pipeline-cleanup bundle Phase 9 / frozen decision 0.11.
+#
+# Phase 5 reconciliation (2026-05-13, architecture diligence audit
+# ARCH-H1): of the 64 entries below, 35 are legitimate wire-protocol
+# carve-outs (SCEP RFC 8894 = 8 entries, ACME RFC 8555 default + per-
+# profile = 27 entries) that MUST stay. The remaining 29 are REST-
+# shaped routes whose OpenAPI ops were deferred during their original
+# Bundle 2 / audit-2026-05-10 / 2026-05-11 work. Burn-down plan:
+#
+#   Sprint A (per-cluster, ~7-8 ops each):
+#     Cluster 1: auth/sessions + auth/oidc (12 ops)
+#     Cluster 2: auth/breakglass + auth/users + auth/runtime-config (8 ops)
+#     Cluster 3: audit/export + demo-residual/cleanup + auth/logout +
+#                auth/breakglass/login + auth/oidc/{login,callback,bcl} (9 ops)
+#
+# Each authored OpenAPI op needs request/response schemas (not
+# placeholders) so the generated client at web/orval.config.ts emits
+# typed signatures. When an op lands, delete the corresponding entry
+# below + bump the openapi-handler-parity.sh expected counts.
 
 documented_exceptions:
   - route: "GET /scep"