fix(hosted): harden versioned local run reporting

[abrichr]

Outcome

Hardens the versioned local run and break reporting rails for the hosted
control plane while preserving the public Flow 1.18.0 caller and CLI surface.

Contract

• Emits exact run-summary/v2 and break-summary/v2 payloads.
• Sends only bounded counters, enums, identifiers, timestamps, and the semantic
  bundle digest; no names, URLs, parameters, screenshots, free-form report
  content, or effect hashes.
• Preserves one run UUID across break and resume.
• Preserves the public report_run(..., deployment_kind=..., org_id=...)
  keyword inputs and report-run --deployment-kind/--org-id flags as
  deprecated compatibility no-ops. Authenticated server identity remains
  authoritative and neither value enters the v2 payload.
• Uses one bearer credential, refuses redirects, caps report input at 8 MiB and
  outbound payloads at 16 KiB.
• Creates .report_run_id race-safely with 0600, exclusive creation, fsync,
  descriptor-based O_NOFOLLOW where available, and descriptor/inode binding
  on every platform before reading.
• Parses both successful rails through one fail-closed response validator:
  canonical run UUID, exact expected status, boolean duplicate when present,
  and canonical halt/teach bindings for break responses. Malformed or
  non-object JSON cannot become emitted=true.
• Local reports are visibility/provenance records, not a billing meter.

Verification

• Exact-head protected CI
  29696104223
  passed every required job, including full unit/coverage, browser E2E, Windows
  mock, Linux AT-SPI, Python compatibility, typing, PHI, and wheel validation.
  Claims run
  29696104222
  also passed.
• Focused hosted tests: 138/138.
• Full non-E2E local regression reached 1,107 passing with zero failures before
  an environment-sensitive ONNX/OCR test stalled; the run was stopped rather
  than allowed to take over the machine indefinitely. The initial sandboxed
  attempt reached 2,270 passing but its localhost-dependent tests were refused
  by the execution sandbox, so neither partial run is represented as a full
  suite pass.
• Repository-pinned Ruff 0.15.21: package lint and full source/test format
  checks pass.
• Focused mypy for both changed source modules passes. Whole-package local mypy
  reports one pre-existing benchmark/dom_arm.py error on exact main; the
  protected PR matrix remains authoritative.
• Consistency, claims, PHI, and diff gates pass.
• Exact-source wheel/sdist built and inspected: MIT boundary intact; no
  openIMIS, AGPL, or repository-only benchmark files.

Release sequencing

Flow 1.18.0 is already public. This PR must therefore release, if accepted,
under the next patch or minor version selected by the semantic-release result;
it does not pre-assign that version.

The completed managed-production qualification used the immutable Flow 1.17.1
/ Modal 1.5.2 report-v2 managed runner. That qualification validates the
managed callback/report path, but it does not directly qualify these local
run-summary/v2 / break-summary/v2 reporting rails. Their evidence is this
PR's paired Cloud contract, adversarial tests, protected CI, and a separately
scoped local-report qualification before release.

[abrichr]

Protected exact-head CI is green. Release remains intentionally held: the first real managed-run compatibility trial against Cloud deploy 6a5cc853 exposed PostgreSQL 42804 in Cloud's apply_run_callback (text status assigned to run_status enum). No workflow action or false success was recorded; the run remains queued and single-flight. Cloud's additive callback repair, deployment, stranded-run reconciliation, and a fresh 3/3 hosted trial matrix will complete before this PR is marked ready or Flow 1.18 is released.