fix(metrics): 14 Snapshot counters reach /stats but never Prometheus, including every summarize/agentdiet failure signal

[OsherElhadad]

proxy/promexport.go is a hand-maintained list of promLine(&b, …, s.<Field>) calls with nothing
asserting that every metrics.Snapshot field is either exported or deliberately exempt. That is how
cg_expand_unresolved_total went unexported and how cg_frozen_decisions_total shipped reading a
permanent 0 (#154).

The coverage test added in #155 reflects over Snapshot's 66 fields and requires each to be exported
or listed with a reason. Fourteen are recorded there as genuine gaps rather than exempted to make
the test pass. Listing them here so they are trackable outside a test file.

Alertable — how you would notice a component silently failing

Field	Why it matters
SummarizeTimeouts, SummarizeErrors	summarize restructures the whole transcript. Silent failure means no compaction and no signal.
AgentDietTimeouts, AgentDietErrors	Same, for the agentdiet reflection pass.
LLMTruncated	A cheap-model reply cut short produces a partial filter. Distinguishable from an error only by this counter.

cg_llm_errors_total and cg_llm_timeouts_total are exported for extract_llm, so the pattern
exists and these three components simply never got the same treatment.

Savings denominators

Field	Why it matters
AttemptedTokens	The divisor for "of what we tried to compact" — one of the four labelled denominators.
FrozenTokens	The cache-frozen ceiling. /metrics cannot express why savings plateaued without it.
SavingsPctAttempted, SavingsPctNewInput	Two of the four headline ratios.

The dashboard makes a point of naming what each denominator divides by; Prometheus currently cannot
compute any of them, because the divisors are not exported.

Remaining

SyncEnforced, CompactionResets, UpstreamMsAvgBypassed, SSETTFBMsAvgBuf,
SSEExpandAfterStream.

Notes for whoever picks this up

• Each needs a judgment about whether it is a counter or a gauge, and a HELP sentence in the
  style the neighbours use — proxy/promexport_help_test.go enforces a convention about the
  cg_*-vs-dashboard disagreement sentence. This is real review surface, not boilerplate, which is
  why fix(dash): report unpriced components in tokens, state the cache-frozen ceiling, and export the counters that only reached /stats #155 did not bulk-add them.
• Several are host-filled: renderMetrics holds the bare h.agg.Snapshot(), and fields the /stats
  handler fills are zero at export time. That is the fix(metrics): cg_frozen_decisions_total reports 0 on every scrape — renderMetrics never fills the Snapshot fields it reads #154 trap — check the source before wiring a
  field, or you will export a constant. Read the counter from its owner the way the cg_llm_* family
  does.
• As each is exported, delete its entry from notExportedWhy. Stale entries fail the test, so the
  list cannot drift out of date in either direction.
• CLAUDE.md:31 claims OpenTelemetry gen_ai.* semantic conventions, which is aspirational —
  all 60 existing metrics are cg_* (chore: CLAUDE.md claims gen_ai.* semconv, but all 60 metrics are cg_* #146). Follow the cg_* convention here; this is not the place
  to start a semconv migration.