Cardano Updates

When `hydra-status-url` is provided, the action connects to the
hydra-github-bridge SSE endpoint for real-time build status updates
instead of polling the GitHub API. This dramatically reduces latency
from minutes (poll interval) to seconds (event-driven).

SSE mode features:
- One-shot check on cached status before opening the stream
- Event filtering by check-run name (ignores unrelated jobs)
- Process substitution to avoid subshell exit trap
- read -t 60 to enforce timeout during idle streams
- Periodic re-checks every 120s as CDN buffering safety net
- Graceful fallback to poll mode on connection failure

Without `hydra-status-url`, the action behaves exactly as before
(pure GitHub API polling with exponential backoff).

Copyright (c) Moritz Angermann <[email protected]>, Input Output Group.
SPDX-License-Identifier: Apache-2.0

When `hydra-status-url` is provided, the action connects to the
hydra-github-bridge SSE endpoint for real-time build status updates
instead of polling the GitHub API. This dramatically reduces latency
from minutes (poll interval) to seconds (event-driven).

SSE mode features:
- One-shot check on cached status before opening the stream
- Event filtering by check-run name (ignores unrelated jobs)
- Process substitution to avoid subshell exit trap
- read -t 60 to enforce timeout during idle streams
- Periodic re-checks every 120s as CDN buffering safety net
- Graceful fallback to poll mode on connection failure

Without `hydra-status-url`, the action behaves exactly as before
(pure GitHub API polling with exponential backoff).

Copyright (c) Moritz Angermann <[email protected]>, Input Output Group.
SPDX-License-Identifier: Apache-2.0

CDN proxies like Cloudflare may buffer SSE events, preventing real-time
delivery of status updates to the client. This caused the script to
hang indefinitely on the SSE stream even after the build completed.

Three fixes:

1. Periodic one-shot re-check: every 120s, poll the bridge's cached
   /status endpoint directly. This catches status changes even when
   SSE events are buffered or lost.

2. Reserve time for fallback: set curl --max-time to TIMEOUT - 120s
   so there's always 2 minutes left for one-shot re-check + polling.

3. Final one-shot re-check after SSE stream ends, before entering
   the polling fallback loop.

Two fixes for SSE mode reliability:

1. Use `read -t 60` instead of blocking `read` so check_timeout fires
   every 60s even when the SSE stream is idle. Without this, if the
   stream goes idle (e.g. Cloudflare buffering, no new events), read
   blocks indefinitely and the script's TIMEOUT is never enforced.

2. After the SSE stream ends, re-check the one-shot endpoint before
   falling back to polling. This catches the case where the target
   status changed while we were connected but the SSE event was lost
   in transit (common with CDN/proxy intermediaries).

The SSE event loop used `curl | while read`, which creates a subshell
for the while loop. This means `exit 0/1` inside the loop only
terminates the subshell — not the script — causing SSE mode to always
fall through to the polling fallback.

Fix by using process substitution (`< <(curl ...)`) so the while loop
runs in the current shell and `exit` works as intended.

The one-shot endpoint returns a map keyed by check-run name, not a
single status object. Extract .[HYDRA_JOB].conclusion instead of
.state from the top level.

Similarly, SSE events carry individual check-runs — filter by .name
matching HYDRA_JOB so we only react to the relevant check-run.

The debug echo inside poll_github() was captured by the command
substitution `conclusion=$(poll_github)`, making `conclusion` a
multi-line string that never matched the `success)` or `failure)`
patterns in the case statement. This caused the action to always
fall through to the `*` wildcard and wait until timeout.

Redirect both debug echo lines to stderr (`>&2`) so only the
actual gh api result is captured in the return value.

- Add timeout input (default: 3600s / 1 hour), exit code 2 on timeout
- Add max-delay input (default: 300s) to cap exponential backoff
- Exponential backoff: start at delay, double each iteration, cap at max-delay
- Fix multi-value conclusion from paginated API results (tail -1)
- Add iteration counter and elapsed time to all log messages
- Use GitHub Actions error annotations for validation/timeout failures

When hydra-status-url is configured, connect to the hydra-github-bridge
SSE endpoint for real-time build status instead of polling the GitHub API.
Falls back to polling automatically on connection failure.

This requires the SSE endpoint to be deployed on hydra-github-bridge.

- Add timeout input (default: 3600s / 1 hour), exit code 2 on timeout
- Add max-delay input (default: 300s) to cap exponential backoff
- Exponential backoff: start at delay, double each iteration, cap at max-delay
- Fix multi-value conclusion from paginated API results (tail -1)
- Add iteration counter and elapsed time to all log messages
- Use GitHub Actions error annotations for validation/timeout failures