Tiered build/test reporting: --report=counts|targets|per-rule|full (pay only for detail materialized)

[paul-hammant]

Motivation

blazerules exposes output-detail tiers (COUNTS → CODES → DECISIONS → BITMASKS) so a caller pays only for the per-record detail it actually materializes — materialization is the cost, so the caller picks the least-expensive level that answers its question.

aeb's analogue is build/test reporting verbosity. Today the [telemetry] block is a single fixed renderer emitted at the end of every build. There's no way to ask for "just the pass/fail totals" (cheap, CI-friendly) versus "per-node cache attribution and timing" (expensive, debugging). This is the same least-expensive-level-required principle, and it directly targets the ◐ Build telemetry and ◐ Test orchestration rows in the scope table.

Proposal

Add --report=<tier> with escalating detail (names illustrative):

Tier	Materializes
counts	aggregate pass/fail, total targets, total wall-time. No per-node rows.
targets	counts + per-target line: label, wall-time, cache hit/miss.
per-rule	targets + per-test-target pass/fail counts (the 14/14 PASS / 2/3 FAIL granularity already parsed from aeocha).
full	per-rule + per-node toolchain-invocation detail (what the current block plus future renderers show).

• Default tier = today's behaviour (whatever the [telemetry] block currently emits), so no regression.
• Renderer stays pure-render over the in-memory telemetry records + .rc markers — same shape as tools/aeb-graph.ae reading _edges.txt. No new I/O sources.
• Tier is a filter on what gets rendered, not what gets collected — records are collected once; the tier decides how much is printed.

Acceptance criteria

• aeb --report=counts <target> prints aggregate-only, no per-node rows.
• aeb --report=full <target> prints everything the current [telemetry] block does, plus per-node detail.
• Default (no flag) is unchanged from today.
• The renderer is pure over telemetry records + markers (no extra probing/exec).
• tests/ string-builder coverage for the tier→render mapping.

Not being asked

Not asking for file-dump or web-view renderers (those are separate build.render_telemetry plug-ins already noted in LLM.md). This is only the CLI verbosity dial over the existing stdout renderer.