fix(mcp): preserve coverage scope labels

[aaiyer]

What does this PR do?

check_index_coverage stored each normalized scope with yyjson's non-copying
string API even though the source was a loop-local buffer. With multiple
scopes, later iterations could overwrite earlier labels; AddressSanitizer also
reported a stack-use-after-scope when the response was serialized.

Use yyjson_mut_obj_add_strcpy so the mutable document owns each scope label.
Add a regression that requests three distinct, equal-length scopes and asserts
that every returned label and its order are preserved.

This PR is intentionally limited to one issue and two files:

• src/mcp/mcp.c
• tests/test_mcp.c

It does not change coverage semantics, tool inputs, schemas, pagination, or
status handling.

Verification

• Reproduce-first: the focused MCP sanitizer suite aborted before the fix with
  stack-use-after-scope in handle_check_index_coverage during yyjson
  serialization.
• make -j4 -f Makefile.cbm test-focused TEST_SUITES=mcp: 166/166 passed.
• scripts/lint.sh --ci with CI's clang-format 20.1.8 and cppcheck 2.20.0:
  passed.
• make -j"$(nproc)" -f Makefile.cbm security: all seven security layers
  passed, including MCP robustness 23/23.
• Focused review: 0 High, 0 Medium, 0 Low findings.

The broad scripts/test.sh run reached the end with all test cases reporting
PASS, then LeakSanitizer exited nonzero for a 4,976-byte allocation originating
from the unchanged cli_install_plan_receipt_no_mutation_issue388 test path
(tests/test_cli.c:2398 / src/cli/cli.c:7539). This PR does not touch that
path; it is being kept out of this one-issue change and investigated separately.

Checklist

• Every commit is signed off (git commit -s) — required, CI rejects
  unsigned commits (DCO, see CONTRIBUTING.md)
• Tests pass locally (make -f Makefile.cbm test) — all cases passed, but
  the separate unchanged CLI LeakSanitizer finding above made the process
  exit nonzero
• Lint passes (make -f Makefile.cbm lint-ci)
• New behavior is covered by a test (reproduce-first for bug fixes)

[DeusData]

Thanks for isolating the yyjson ownership defect and retaining a multi-scope sanitizer regression. Because this is a stack-use-after-scope in an MCP response path, I’ve classified it as a high-priority 0.9.1-rc memory-safety fix. Review will stay focused on the copying API, response lifetime, and whether equivalent loop-local yyjson strings exist in adjacent coverage output.