feat(security): SEC009 language-aware dispatch (PR #12)

[wei9072]

Summary

PR #10 fixed multi-language coverage for SEC010 (weak RNG); Round 9
validation revealed SEC009 (weak hash) had the same coverage
gap. The old matcher (`last == "md5" | "sha1" | ...`) only
fired on Python `hashlib.md5(...)`. Real-world Go and Java auth
code went uncaught:

```go
import "crypto/md5"
func HashPassword(password string) string {
h := md5.Sum([]byte(password)) // <-- silent before PR #12
...
}
```

```java
import java.security.MessageDigest;
public static String hashPassword(String password) throws Exception {
MessageDigest md = MessageDigest.getInstance("MD5"); // <-- silent before
...
}
```

This PR mirrors PR #10's per-language dispatch architecture for
SEC009.

Per-language matchers

Language	Match shape	Notes
Python	`hashlib.md5` / `hashlib.sha1`	unchanged
Node / JS	`crypto.createHash('md5'|'sha1')`	algo in first string arg
Go	`md5.Sum` / `md5.New` / `sha1.X`	Layer 1 resolves `crypto/md5` vs same name from elsewhere
Java / Kotlin	`MessageDigest.getInstance("MD5"|"SHA-1"|"SHA1")` + Apache `DigestUtils.md5Hex`/`sha1Hex`	algo in string arg
C#	`MD5.Create` / `SHA1.Create` / `MD5CryptoServiceProvider` / `MD5Managed`	receiver-anchored to avoid `SHA1024` collisions
PHP	global `md5`/`sha1` + `hash('md5'|'sha1', ...)`	algo in string arg for `hash()`

New helper `first_arg_is_weak_alg_string()` inspects the call's
first string-literal argument and matches against `md5` / `sha1` /
`sha-1` (case-insensitive). Used by Node, Java, and PHP.

enclosing_security_context fix (PR #11 parallel)

Same bugs PR #11 fixed in `enclosing_token_context`:

• Multi-language assignment node kinds added (Go
  `short_var_declaration`, Java `local_variable_declaration`, etc.)
• Function-name needle check at function-shape level. Round 9
  production case: `func HashPassword(password string)` carries the
  `password` needle through its name even when the local
  assignment uses generic `h := ...`.
• Walks past inner blocks (loop bodies / if bodies don't terminate
  the search).

Tests

8 new SEC009 tests covering each language's positive and negative
cases. Plus the `md5_for_etag_does_not_block` Python negative is
preserved (etag context must stay silent).

`cargo test --workspace`: 164 / 164 pass (was 156; +8 new).

Live MCP confirmation

Round 9's starting-go/auth.go before this PR:
```
FINDINGS: 19 total (security=2)
[security] SEC012 @line 17 ...
[security] SEC010 @line 26 ...
```

After PR #12:
```
FINDINGS: 20 total (security=3)
[security] SEC009 @line 11 weak hash ... (`password`)
[security] SEC012 @line 17 ...
[security] SEC010 @line 26 ...
```

Round 9's starting-java/Auth.java similarly goes from 1 → 3 SEC
findings.

Test plan

• `cargo test --workspace` — 164 / 164 pass
• Live MCP fires SEC009 on Round 9 Go + Java starting code
• Existing SEC009 etag-context negative still passes
• PR fix(security): SEC010 FP on secrets.choice / os.urandom #9, feat(aegis-core): Layer 1 per-file fact cache + SEC010 multi-language dispatch #10, fix(security): SEC010 reads function name + walks past inner blocks #11 regression tests still pass
• CI green on push

🤖 Generated with Claude Code