fix(langfuse): review fixes 3 - branch-quiet notify, 1h cache-write price, lean scorer, trigger label, memory byte budget

[aaronjmars]

Fixes from the 2026-10-16 Langfuse review of aeon-agent and miroshark-aeon (follow-up to #1230).

1. No real notifications from test branches. "Send pending notifications" now needs github.ref_name == github.event.repository.default_branch, like the health-issue step fix(langfuse): branch-gated health issues, traced feed render, prices, usage-row timing #1230 gated. workflow_dispatch runs of holdings on verify/langfuse-models / verify/langfuse-slicing delivered 4 real Telegram messages ("Delivered 1/1" in aeon-agent runs 38778704213, 38789847709 and miroshark-aeon runs 38779134169, 38789854033). Repo variable NOTIFY_ON_BRANCH=1 opts a test branch back in. Skipped payloads live in $RUNNER_TEMP, so nothing is redelivered later. Documented in docs/CONFIGURATION.md.
2. 1-hour cache writes priced at the 1-hour rate. Each usage <model> row now also sends langfuse.observation.usage_details with the Langfuse price keys (input, output, input_cache_read, input_cache_creation_1h, input_cache_creation_5m). Langfuse takes that attribute over gen_ai.usage.*, which has one cache-write count it priced at the 5-minute rate (1.25x input). Claude Code on a subscription writes 1h caches (2x), so cost read ~35% low on every Sonnet run (the reported aeon.cost_usd_reported fit $4/M writes on 7 of 7 runs). The split comes from the envelope's usage.cache_creation.ephemeral_1h_input_tokens / ephemeral_5m_input_tokens (checked in a real Claude Code 2.1.301 --output-format json result; modelUsage has no split, so with more than one model each gets the run's 1h share). No split in the envelope = one input_cache_creation count, as before. gen_ai.usage.* stays for other OTLP backends. fleet-scorecard.mjs prices cache writes at 2x input too (token-usage.csv has no TTL column).
3. Lean scorer. New run-harness --no-tools (claude: --tools ""; other harnesses ignore it). The claude scorer uses it and runs from an empty scratch dir, the way fix(langfuse): branch-gated health issues, traced feed render, prices, usage-row timing #1230 runs the feed render, so a 1-turn text grade no longer loads tool schemas and the repo CLAUDE.md (~18.7k read + ~18k write per call on both instances). Other harnesses keep their current call.
4. Trigger label tells cron from manual. Every skill starts as workflow_dispatch, so trigger was workflow_dispatch for cron, chain and manual runs alike. New optional aeon.yml input trigger: the Scheduler passes cron, chain-runner chain; langfuse-otel.sh maps them to cron / chain / manual (only those values, the input is free text). Both dispatchers retry once without the field when an older aeon.yml answers "Unexpected inputs", so a fork that syncs the Scheduler before its aeon.yml keeps running.
5. MEMORY.md byte budget. aeon-agent's MEMORY.md is 88 lines but 27.7 KB and loads into every run; the cap was lines only. memory-flush now holds it to 8 KB with no line over 300 chars (step 3e), and memory_prep.py window prints the measured size plus the longest lines so the model trims by numbers. CLAUDE.md (and AGENTS.md) say the same.

Verified:

• A real usage_details row sent to the aeon-agent Langfuse project (environment fix-test): 100k 1h writes priced $0.40 (input_cache_creation_1h), total $0.422 for the row.
• claude -p --output-format json envelope fields checked locally; run-harness claude --no-tools from a scratch dir returns a valid envelope (48.5k -> 19.6k context tokens on a laptop with a large global CLAUDE.md; on a runner the drop is to the bare system prompt).
• Local: test_otel_span.sh (1h/5m split per model + no-split fallback), test_langfuse_otel.sh (cron / chain / manual / free-text / non-dispatch), test_memory_prep.py (size report), test_fleet_scorecard.mjs, test_chain_runner*.sh, the full ci-tests list (only test_chain_runner.sh fails on macOS, same on main: BSD grep has no -P), validate-config.js, gen-agents-md.js --check, actionlint, shellcheck error gate.

Instances: aeon-agent and miroshark-aeon get the aeon.yml hunks for 1 and the trigger input now (instance PRs), so the Monday aeon-update brings the Scheduler / chain-runner / script changes onto a workflow that already has the input. The scorer hunk (3) needs the new run-harness flag, so it arrives with that sync.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}