fix(overseer): tolerate trailing text after a health-review decision's JSON

[rysweet]

Summary

The Overseer's agentic HEALTH-REVIEW rail ([standing] goal) reads the
journal, reasons to a small set of typed DECISION markers
(LAUNCH_RECIPE=/ESCALATE_GOAL=/HEALTH_REVIEW_COMPLETE=), and the thin Rust
rail (src/overseer/health_review.rs) routes each onto the same gated
capability path every other fix uses. A prior hardening made the parser
tolerant of leading markdown decoration, but each decision's JSON payload was
still handed straight to serde_json::from_str, which rejects any trailing
non-whitespace after the object.

So a well-formed decision an agent wrote as
LAUNCH_RECIPE={...} — fixes the crash-loop parsed to zero interventions
while the terminal HEALTH_REVIEW_COMPLETE= marker still parsed "successfully" —
the degraded-pass escalation ladder never fired and a real crash-loop's
remediation was lost with no signal. That is the exact silent-drop the
leading-decoration fix closed, but on the trailing side, and it directly defeats
"DRIVE remediation through EXISTING capabilities".

Fix (thin rail, mechanical, fail-closed)

Add extract_leading_json_object, a balanced-brace scanner that respects JSON
strings + escapes, used by parse_launch_decision/parse_escalate_decision
before serde. It:

• extracts the leading {…} object and discards any trailing clause (trailing
  text never leaks into the parsed brief / operator-facing fields);
• returns None for a payload that does not start with { or whose braces never
  balance, so the caller falls back to the raw payload — which serde then rejects
  — rather than half-parsing a truncated/broken decision (fail-closed);
• never edits the JSON it returns (serde still validates every required field);
• returns the first object on a concatenation, so a trailing {...} cannot
  smuggle a second decision.

Byte-identical parse for exact-JSON payloads. No prompt / capability /
gate / output-contract change. Critically, no new imperative failure-plumbing
— no record_step_failure at an origin site, no consecutive-failure counter, no
N-identical-failure threshold. Reasoning stays broad; authorization stays narrow;
the anti-pattern the standing goal forbids is not touched.

Changed surfaces (all internal)

File	Change
src/overseer/health_review.rs	new extract_leading_json_object + wired into both parse fns + doc bullet + 6 unit tests
tests/gadugi/overseer-health-review.sh	pin the 6 new tests as required invariants
tests/qa-scenarios/overseer-health-review-trailing-text.yaml	new outside-in QA scenario
docs/concepts/overseer-agentic-health-review.md	document the trailing-side symmetry

No user-facing surface changed (the marker contract, prompt, and operator
notification text are unchanged); this is a parser-robustness fix on an internal
rail, plus the concept doc that already documents that rail.

---

Merge-ready evidence

1. QA / gadugi scenarios

• New QA scenario tests/qa-scenarios/overseer-health-review-trailing-text.yaml
  drives the hermetic unit tests + full overseer suite (YAML validated: 5 keys, 6 steps).
• Gadugi scenario tests/gadugi/overseer-health-review.sh pins the 6 new
  tests via require_test and was run green:

  == (a) recipe-runner json mode surfaces the health-review markers ==
  OK: the recipe-runner seam forwards the health-review decision markers verbatim.
  == (b) in-tree overseer health-review tests ==
  health-review tests passed: 54
  PASS: overseer-health-review scenario ([standing])
  

2. Docs

docs/concepts/overseer-agentic-health-review.md updated with the trailing-side
symmetry paragraph (+ last_updated). All other changed surfaces are
internal-only (parser rail + test harness).

3. Quality-audit (3 SEEK→VALIDATE→FIX cycles, ended clean)

• Cycle 1 — correctness. Brace-depth scanner: cannot underflow (payload must
  start with {, returns at depth==0), correct UTF-8 slicing (i + len_utf8),
  string/escape state respected. VALIDATED by the 6 new tests incl.
  extract_leading_json_object_basics. No fix needed.
• Cycle 2 — security / XPIA / fail-closed. Cannot smuggle a second decision
  (returns the FIRST object; trailing discarded, never re-parsed as a marker);
  trailing text never leaks into task_description/next_step; unbalanced →
  None → serde reject → skipped. VALIDATED by brace-in-string + unbalanced +
  existing 54 tests. No fix needed.
• Cycle 3 — regression / philosophy. Zero new imperative failure-plumbing;
  no print!/println!; no Bridge naming; byte-identical exact-JSON parse;
  require_test names match test fns exactly. VALIDATED by fmt + clippy
  --all-targets --all-features -D warnings + full health_review filter across
  all targets. Clean final cycle — zero critical/high, zero medium
  correctness/security findings.

4. CI-relevant gates (run locally, all green)

• cargo fmt --all -- --check → clean
• cargo clippy --release --no-deps -- -D warnings (pre-commit) → clean
• cargo clippy --all-targets --all-features --locked -- -D warnings → clean
• cargo test --lib overseer::health_review::tests → 35 passed, 0 failed
  (incl. the 6 new tests)
• cargo test --lib overseer:: → 749 passed, 0 failed
• cargo test --all-features --locked health_review (all targets) → ok, 0 failed
• pre-push race-subset test → 482 passed, 0 failed

6. Focused diff

4 files, +215/−7, all within the health-review rail + its test harness + its
concept doc. No unrelated edits.