[log-classifier] Remove generic OSDC rule, use Bedrock inference profiles, widen LLM context (#8391)

## Summary

Three related fixes to the log-classifier Lambda:

1. **Remove the generic `OSDC step script failure` rule.** `[OSDC] Step
script exited...` is a generic wrapper message that hides the real
failure, so classifying on it produced useless results. Removing it lets
these logs fall through to the generic `GHA error` catch-all (and then
the LLM).

2. **Keep `GHA error` as the lowest-priority rule.** Rules in
`ruleset.toml` are ordered by priority (first = highest), and `main.rs`
only kicks off the LLM classifier when the best match is the *last*
rule. So the generic `^##[error](.*)` catch-all must stay last — `New
modules are not documented correctly` is moved above it.

3. **Use `us.*` cross-region inference profile IDs for Bedrock.** The
bare `anthropic.claude-*` on-demand IDs are not directly invokable for
these models; the `us.anthropic.*` inference profiles are required.

4. **Widen the LLM error-context window from 100 → 500 lines.**
`make_query` centers the snippet on the matched line: `[error_line -
num_lines/2, error_line + num_lines/2]`. The catch-all match often lands
on a wrapper line that sits well above the real cause, so a narrow
window misses it. See the 90075305031 case below: the match is at line
600 but the root cause is at line 427 (173 lines up) — with 100 the
window is only `[550, 650]` and the LLM settles for a nearby symptom
line; with 500 (`[350, 850]`) it reaches the real error.

   | context | surfaced line for 90075305031 | correct? |
   |---|---|---|
| 100 | line 453: `##[error]TypeError: Cannot read properties of null
(reading 'jobPod')` (a symptom/boilerplate line) | no |
| 500 | line 427: `##[error]Error: pod failed to come online ... backoff
timeout` | yes |

Caveat: 500 is empirically fitted (`±250` lines), not principled — a log
whose real cause is even further from the wrapper line would still be
missed. A more robust fix would anchor the snippet on the last error
region rather than widening symmetrically, but that is a larger change.
Cost/latency of ~5x prompt tokens per fallback call is negligible for
Haiku and the fallback only fires on the catch-all rule.

## Relationship to #8382

This is an alternative to @jeanschmidt's #8382, which drops the
CI-wrapper boilerplate lines at preprocessing time. This PR reaches the
same outcome via a different mechanism: every log falls through to the
`GHA error` catch-all and the **LLM fallback** re-locates the meaningful
error line.

## Test plan

- `cargo check` passes (only pre-existing deprecation warnings,
unrelated to these changes).

Re-ran the four real prod jobs from [Jean's validation
table](#8382 (comment))
through this branch (read-only, no DynamoDB write):

| job_id | Jean's AFTER (#8382) | This branch's FINAL | verdict |
|---|---|---|---|
| 90151654008 (P1 exit-code) | `Action failed... non-human actor:
pytorch-bot` | `##[error]Action failed with error: Workflow initiated by
non-human actor: pytorch-bot (type: Bot). Add bot to allowed_bots
list...` (line 472, Bedrock haiku) | same |
| 90075305031 (P2 OSDC) | `Error: pod failed to come online ... backoff
timeout` | `##[error]Error: pod failed to come online ... is unhealthy
with phase status Pending: backoff timeout` (line 427, Bedrock haiku) |
same |
| 90103895526 (P3 container) | `[OSDC] Step cancelled by GitHub Actions
(received SIGINT)` |
`inductor/test_cpu_repro.py::CPUReproTests::test_vec_compare_op_cpu_only
Command took >60min, returning 124` (line 9532, Bedrock haiku) | better
— surfaced the hung test that caused the SIGINT |
| 90131159825 (P3 container) | `Error: pod failed to come online ...
backoff timeout` | `Pod lf-l-... phase=Pending ... waiting=[job:
ErrImagePull] (2596s/7200s)` (line 220, Bedrock haiku) | better —
pinpointed `ErrImagePull` |
| P4 jobPod | n/a (0 prod rows in 120d) | not run — no real log exists |
matches Jean's note |

All four surface the real failure instead of the useless `exit code 1` /
OSDC boilerplate; two cases trace past the generic pod/SIGINT symptom to
the actual root cause.

Author: huydhn

aws/lambda/log-classifier/ruleset.toml

@@ -314,13 +314,9 @@ name = 'ghstack merge check error'
 pattern = '^Mergeability Error: .*'
 
 [[rule]]
-name = 'OSDC step script failure'
-pattern = '\[OSDC\] Step script exited.*'
+name = 'New modules are not documented correctly'
+pattern = 'You added the following module\(s\) to the PyTorch namespace .* but they'
 
 [[rule]]
 name = 'GHA error'
 pattern = '^##\[error\](.*)'
-
-[[rule]]
-name = 'New modules are not documented correctly'
-pattern = 'You added the following module\(s\) to the PyTorch namespace .* but they'

aws/lambda/log-classifier/src/bedrock.rs

@@ -119,8 +119,8 @@ pub async fn make_query(
     error_line: &usize,
     num_lines: usize,
 ) -> Option<SerializedMatch> {
-    let model_id_primary = String::from("anthropic.claude-haiku-4-5-20251001-v1:0");
-    let model_id_secondary = String::from("anthropic.claude-sonnet-4-6");
+    let model_id_primary = String::from("us.anthropic.claude-haiku-4-5-20251001-v1:0");
+    let model_id_secondary = String::from("us.anthropic.claude-sonnet-4-6");
     let line_numbers = vec![*error_line];
 
     let log_snippet = get_snippets(log, line_numbers, num_lines / 2, num_lines + 1);

aws/lambda/log-classifier/src/main.rs

@@ -51,7 +51,7 @@ async fn handle(
             // check if match has the lowest priority in the ruleset
             if best_match.rule.name == ruleset.rules.last().unwrap().name {
                 // kick off the llm to get the rule
-                match make_query(&log, &best_match.line_number, 100).await {
+                match make_query(&log, &best_match.line_number, 500).await {
                     Some(llm_match_json) => {
                         body = serde_json::to_string_pretty(&llm_match_json)?;
                         match_json = llm_match_json;