Redact `_block` reasons before they are persisted, like the agent-message path

Problem

ControllerEngine._block (src/autoforge/engine.py, _block) writes its reason to state.block_reason as given. Several of its ~40 call sites embed external text in that reason, most often a GitHubError message (f"... cannot be read: {exc}"), which is gh stderr. The two agent-message paths that also set block_reason (engine.py around lines 1587 and 1772) pass the text through redact() first, with the comment "same redaction boundary as the run log", so persistence is currently redacted for agent text but not for controller-composed reasons.

Reproduced in tests/test_cli.py::test_run_terminal_line_redacts_a_block_reason_the_engine_persisted_raw (added on PR #97): with FakeGitHub.get_pr_error = GitHubError("HTTP 401: Bad credentials (Authorization: Bearer ghp_...)") and a state referencing a PR, resume blocks at ANALYZE_EXECUTE entry and state.json holds the token in the clear. The CLI output (_finish terminal line, status, status --json) is redacted, so nothing is printed, but the file on disk is not.

Why not in PR #97

PR #97 closes #6 / #67, whose scope is what the CLI prints from state; it deliberately never rewrites state.json, and its docs paragraph states that persisted state is not a redaction boundary of its own. Redacting at the persistence side is a separate, engine-level change with its own tests, and state.json is local, untracked, per-operator runtime state, so it is not a leak in the sense of #6. It is still inconsistent with the two agent-message writers and with docs/agent-guides/secrets-and-logging.md.

Recommended direction

Redact once in _block (state.block_reason = redact(reason), and the StepOutcome.message it returns), rather than at each call site, so a new call site is covered by construction. Consider whether GitHubClient should redact error text at construction instead, which would also cover VerificationError paths that quote gh output and reach verification_failures. Keep the CLI output-side redaction from PR #97 as defense in depth; do not remove it.

Relevant files

• src/autoforge/engine.py: _block, _loop_block_reason, the two redact(...) agent-message writers
• src/autoforge/github.py: GitHubError construction sites
• docs/agent-guides/secrets-and-logging.md
• tests/test_cli.py::test_run_terminal_line_redacts_a_block_reason_the_engine_persisted_raw (asserts the current raw persistence; flip it when this lands)

Acceptance criteria

• A _block reason that embeds a GitHubError carrying a baseline-pattern secret is persisted redacted in state.json and reported redacted in the StepOutcome and run log.
• verification_failures entries that quote gh output are redacted before persistence.
• The CLI output-side redaction from PR #97 remains in place and tested.
• docs/agent-guides/secrets-and-logging.md says which writers redact before persistence.

Start with _block in src/autoforge/engine.py, then inspect the agent-message redaction paths and _loop_block_reason. Run tests/test_cli.py::test_run_terminal_line_redacts_a_block_reason_the_engine_persisted_raw, and review GitHubError construction in src/autoforge/github.py. Done means controller and verification-failure reasons are redacted in state.json, StepOutcome, and logs, while CLI tests remain passing and the secrets documentation names the persistence writers.