Add provider-specific OPTIONAL fields to usage.jsonl + drop spec-requires framing

[riddim-developer-bot]

Summary

Previously the backfilled usage.jsonl dropped three high-signal cost fields:

• Codex per-window quota readouts (the most visceral cost-as-it-happens metric)
• Codex reasoning-vs-visible output split
• Claude cache-tier split (5m vs 1h ephemeral cache writes)

The cost-telemetry spec allows OPTIONAL field additions without bumping schemaVersion (per its own §6.3), so this PR adds them. The bake path is now lossless for every signal the cost analysis actually uses.

Spec additions (specs/symphony-cost-telemetry-extension/SPEC.md)

§5.2.1 Input-Token Breakdown

• inputUncachedTokens, inputCachedReadTokens, inputCacheWriteTokens
• inputCacheWriteEphemeral5mTokens, inputCacheWriteEphemeral1hTokens (Anthropic-only; other writers MUST omit)

§5.2.2 Output-Token Breakdown

• outputVisibleTokens, outputReasoningTokens (reasoning is Codex/o-series-only)

§5.2.3 Quota Sample

"quota": {
  "planType": "pro",
  "windows": [
    { "label": "primary",   "windowMinutes": 300,   "usedPercent": 64, "resetsAt": 1779863673 },
    { "label": "secondary", "windowMinutes": 10080, "usedPercent": 57, "resetsAt": 1780187884 }
  ]
}

Generic shape — Codex emits primary (5h) and secondary (7d), but any provider with any number of rate-limit windows fits the structure.

§5.3 Semantics adds SHOULD-sum relations across the breakdown buckets.

Implementation

• transcripts/codex.mjs emits per-turn quota samples in the spec's windows shape directly (no lossy reshape).
• transcript-to-usage.mjs emits every breakdown field the transcript carries plus the quota object.
• usage-aggregator.mjs prefers the breakdown fields when present; falls back to the REQUIRED inputTokens/outputTokens totals when not.
• bin/llm-cost.mjs quota printer is now generic over windows (renders every label the provider reports, not hard-coded to primary/secondary).

Framing fixes folded in

The broken "Symphony Telemetry Extension Spec" framing for the workspace convention was still on main (the per-issue-workspace requirement is in OpenAI Symphony's parent SPEC.md §4.1.4 — there is no extension spec for it). Re-applies the correction that got lost when its prior worktree was torn down:

• DEFAULT_CWD_PATTERN broadened to match both spec-default <system-temp>/symphony_workspaces/<ID> and the common in-repo <repo>/.symphony/workspaces/<ID>.
• Issue-ID character class widened from [A-Z]+-\d+ (Linear-specific) to [A-Za-z0-9._-]+ (matches the spec's workspace_key sanitization rule).

README prose changes (root README + package README) — per the user's explicit guidance:

• No prose claims that llm-cost-attribution REQUIRES any extension spec.
• The usage.jsonl bake feature is presented as built-in; spec interop with other tools is mentioned only as an optional side-benefit.
• The OG Symphony spec (https://github.com/openai/symphony/blob/main/SPEC.md) is cited correctly for the per-issue-workspace convention it actually requires.

End-to-end verification on real data

Re-backfilled the full 4,309 sessions / 5 GB of transcripts on this machine:

	Before this PR	After this PR
usage.jsonl size	83 MB	125 MB (40× still smaller than the 5 GB source)
llm-cost EPAC-1940 --from-usage output matches transcript-source?	partial (lost quota / cache split / reasoning)	identical
Query time	0.3s	0.3s

Sample backfilled Codex record now includes:

{
  ...
  "inputUncachedTokens": 19975,
  "inputCachedReadTokens": 3456,
  "outputVisibleTokens": 135,
  "outputReasoningTokens": 60,
  "quota": {
    "planType": "pro",
    "windows": [
      { "label": "primary",   "windowMinutes": 300,   "usedPercent": 6, "resetsAt": 1778021087 },
      { "label": "secondary", "windowMinutes": 10080, "usedPercent": 7, "resetsAt": 1778548110 }
    ]
  }
}

Sample backfilled Claude record with cache writes:

{
  ...
  "inputUncachedTokens": 3,
  "inputCachedReadTokens": 0,
  "inputCacheWriteTokens": 45728,
  "inputCacheWriteEphemeral5mTokens": 45728
}

Test plan

• 33 of 33 tests pass (was 32) via node --test. Added: Symphony-spec-default cwd test, breakdown field tests, quota round-trip test. Updated: existing aggregator quota fixtures to the new spec shape.
• node --check clean on every .mjs (including the new src/quota.mjs)
• Backfilled 190,481 records and validated every one passes validateUsageRecord
• llm-cost EPAC-1940 (transcripts) and llm-cost EPAC-1940 --from-usage <file> produce identical token totals, turn counts, model lists, wall-clock spans, AND quota readouts