fix(security): SEC010 reads function name + walks past inner blocks

[wei9072]

Summary

Round 9's production validation (gpt-5.4-mini × Go/Java brownfield)
caught two related bugs in `enclosing_token_context` that hid
SEC010 from real Java code, even though my unit tests passed:

The Java production case that didn't fire:

```java
public static String generateSessionToken() {
String chars = "abcdef0123456789";
StringBuilder token = new StringBuilder();
for (int i = 0; i < 16; i++) {
int idx = new Random().nextInt(chars.length()); // ← weak, but silent
token.append(chars.charAt(idx));
}
return token.toString();
}
```

Walking up from `new Random().nextInt(...)`:

• variable_declarator (`idx = new Random().nextInt(...)`) — needle check on "idx" fails
• local_variable_declaration — same
• block (for body) — opaque-id heuristic on body text fails (no `string.ascii_letters`); previous code `break`d here
• for_statement
• block (method body)
• method_declaration — name is `generateSessionToken`, contains `session` AND `token`, but we never reached it

Three fixes

1. Don't `break` at inner blocks. Only function-shape nodes
   terminate the upward walk. Inner `block` / `function_body` nodes
   contribute text to the opaque-id heuristic, then keep climbing.
2. Read function name field. When the walk lands at a
   function-shape node, run the same needle list against its
   `name` field. `generateSessionToken` matches `session` /
   `token`. `make_short_code` matches `short_code`.
3. Add `method_declaration` + `constructor_declaration` to
   the function-shape match list. Java tree-sitter uses these,
   not `method_definition`.

Tests

• `sec010_java_new_random_in_session_token_function_blocks` —
  the Round 9 production case. Was the FN that triggered this PR.
• `sec010_python_random_in_token_function_blocks` — Python
  parallel using `make_session_token` to verify the same fix
  benefits Python.
• `sec010_random_in_dice_function_does_not_block` — function
  name `roll_dice` has no needle, opaque-id heuristic doesn't
  fire either, must stay silent.

`cargo test --workspace`: 156 / 156 pass (was 153; +3 new).

Live MCP confirmation

Pre-fix:
```
$ aegis_validate.py starting-java/Auth.java
FINDINGS ...: 18 total (security=1, signal=16, workspace=1)
[security] SEC012 @line 20 ...
```

Post-fix:
```
$ aegis_validate.py starting-java/Auth.java
FINDINGS ...: 19 total (security=2, signal=16, workspace=1)
[security] SEC012 @line 20 ...
[security] SEC010 @line 28 weak RNG used for security token ...
```

Test plan

• `cargo test --workspace` — 156 / 156 pass
• Live MCP fires SEC010 on Round 9's starting-java/Auth.java
• PR fix(security): SEC010 FP on secrets.choice / os.urandom #9 / PR feat(aegis-core): Layer 1 per-file fact cache + SEC010 multi-language dispatch #10 regression tests still pass (`secrets.choice`,
  `os.urandom`, `SecureRandom`, `crypto/rand` all still silent)
• FP control: `roll_dice` / non-security function names stay
  silent
• CI green on push

🤖 Generated with Claude Code