Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In @.github/workflows/e2e-advisor.yaml:
- Around line 194-198: Guard the node invocation of e2e-advisor/dispatch.mts by
first checking whether the file e2e-advisor/dispatch.mts exists and, if not,
skip execution while writing a deterministic "skipped" artifact to the same
result path (artifacts/e2e-advisor/e2e-advisor-final-result.json) so downstream
steps see a valid JSON result; update the GitHub Actions step that runs node
--experimental-strip-types "$ADVISOR_DIR/tools/e2e-advisor/dispatch.mts" to
perform an existence check and either run dispatch.mts (with the same --result,
--workflow, --workflow-path, --out-dir args) or create the skipped JSON artifact
(containing a stable "skipped" status) and echo/log a clear message so PR runs
do not hard-fail when main lacks dispatch.mts.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@tools/e2e-advisor/comment.mts`:
- Around line 99-107: The parseArgs function currently always consumes argv[i+1]
as a flag value which causes the next flag to be swallowed if no value is
provided; update parseArgs (inside the loop handling arg, key, and parsed) to
check that argv[i+1] exists and does not startWith("--") before assigning
parsed[key] = argv[i+1] and incrementing i; if the next token is missing or is
another flag, set parsed[key] to a boolean true (or undefined per project
convention) and do not increment i so flags are not consumed incorrectly.
- Around line 182-184: The function findExistingComment only requests the first
page of comments (per_page=100) so it can miss the marker on busy PRs; modify
findExistingComment to paginate through GitHub comments by adding a page
parameter (e.g., page=1,2,...) to the github call and loop requesting pages
until either a comment containing marker is found or the returned page is empty
(or fewer than per_page entries), returning the found comment immediately to
avoid extra requests; ensure you keep using the same token and per_page=100 for
efficiency and stop the loop when no more pages remain.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@tools/e2e-advisor/pi-analyze.mjs`:
- Around line 25-27: The environment parsing for PI_E2E_ADVISOR_TIMEOUT_MS and
PI_E2E_ADVISOR_HEARTBEAT_MS is brittle: validate and sanitize values before
using them with setTimeout/setInterval by parsing into integers, ensuring they
are finite and non-negative, and falling back to safe defaults when invalid;
update the parsing logic that produces timeoutMs and heartbeatMs (and any
similar parsing later in the file around the other timer usage) to clamp
negative values to a minimum (e.g., 0 or a sane minimum heartbeat), use
Number.isFinite to reject NaN/Infinity, and then pass only the validated numbers
to setTimeout/setInterval to avoid immediate timeouts or busy loops.
- Around line 126-157: The subprocess stdout/stderr buffers are unbounded
(variables stdout, stderr) and can exhaust memory on noisy runs; limit each
buffer by introducing a MAX_BUFFER_BYTES constant and when appending in
child.stdout.on("data")/child.stderr.on("data") only keep the most recent bytes
up to that limit (e.g., append chunk then, if buffer.length > MAX_BUFFER_BYTES,
truncate to the tail or drop the oldest bytes) and optionally track a counter of
dropped bytes to log later; apply the same logic for both stdout and stderr and
ensure any log or error messages include the fact that output was truncated.