fix: populate AI credits footer for copilot-sdk workflows

[Copilot]

copilot-sdk: true workflows (e.g. pr-triage-agent) were missing agent AIC in the footer — showing only threat-detection AIC (⌖ 38.8 AIC) with no agent cost.

Root cause: Copilot SDK inference bypasses the AWF API proxy; only ambient-context calls hit the proxy, and those carry no model name. Every token-usage.jsonl entry has model: "" → totalAIC = 0 → parse-mcp-gateway never sets the aic step output → GH_AW_AIC is empty downstream. The actual token data is available in the session.shutdown SDK event (data.modelMetrics, data.currentModel), but it was being silently dropped.

Changes

• copilot_sdk_session.cjs — add session.shutdown case to capture modelMetrics/currentModel in a closure variable. Swap finally order to disconnect → write shutdown entry → stream.end so the event fires (synchronous Node.js EventEmitter semantics) while the stream is still writable.

• parse_copilot_log.cjs (normalizeCopilotSdkEventsToTrace) — handle session.shutdown entries: attach modelMetrics to the result entry; use currentModel as the init model instead of the hardcoded "copilot-sdk".

• log_parser_bootstrap.cjs — after parsing, if result.modelMetrics is present, compute AIC via computeInferenceAIC (provider "github-copilot") and export as GH_AW_AIC via core.exportVariable.

• parse_mcp_gateway_log.cjs (writeStepSummaryWithTokenUsage) — when totalAIC === 0, fall back to process.env.GH_AW_AIC set by the preceding bootstrap step and call setOutput("aic", ...) to propagate it as the job-level output consumed by downstream jobs.

The step ordering in the agent job already makes this work: log_parser_bootstrap runs before parse-mcp-gateway, so core.exportVariable in the first step is visible as process.env.GH_AW_AIC in the second.

[pelikhan]

@copilot ensure all messages are logged to stderr as JSONL

[Copilot]

@copilot ensure all messages are logged to stderr as JSONL

Done in 00ae8f1. Replaced the two core.info() / coreObj.info() diagnostic calls I introduced with process.stderr.write(JSON.stringify({ level: "info", message: "..." }) + "\n") in log_parser_bootstrap.cjs and parse_mcp_gateway_log.cjs. The session.shutdown entry written in copilot_sdk_session.cjs was already using JSONL stdout.