Background

Per the symbol-bloat analysis of the latest ESP32-S3 Blink build (fbuild, post #2968):

• Total flash: 330,054 B across 3,428 live symbols
• 8 distinct fl::* functions land in the 1-4 KB band, totaling 13,738 B
• ESP-IDF / libc cluster (~28 KB libc printf, ~10 KB ESP-IDF drivers, ~5 KB Arduino HAL) dominates the absolute top, but those are vendor-controlled. The 1-4 KB band is where FastLED has real leverage.

The smoking-gun pattern

Every single 1-4 KB FastLED symbol is dominated by FL_WARN/FL_ERROR/FL_INFO machinery. Verified across all 8:

These callees are referenced (so they each cost their full size once). But the per-call-site inline cost is what bloats individual functions:

Per FL_WARN/FL_ERROR call site:
  ~200-300 B inline (sstream ctor + format << chain + log_emit call)
  ~100-200 B exception landing pad metadata (string/sstream dtors can throw)
A function with 3-5 log sites ends up with ~1-2 KB of pure logging overhead.

Constraint

Logging must stay enabled — FASTLED_LOG_VERBOSITY=1 (the default for debug + when SKETCH_HAS_LARGE_MEMORY) must continue to produce full log output. Sub-issues that try to no-op log call sites are off-table.

What IS in scope: make the logging infrastructure itself cheaper. Examples:

• Mark string/sstream destructors noexcept -> eliminate landing pads (saves ~100-200 B per call site)
• Push more inline boilerplate into log_emit -> shrink the per-call-site burst
• Consolidate multiple log sites in one function through expected<>::failure(...) carrying the error info to a single bottom-of-function log
• Specialize basic_string hot path so writes from sstream don't pay the variant-storage dispatch
• Extract validation/error paths into [[gnu::cold]] helpers (out-of-line, doesn't bloat the caller)
• Convert dynamic-driver dispatch (which currently inlines multiple writer templates) to virtual dispatch

Sub-issues (will be linked as filed)

Cross-cutting (affects all 8 functions):

• bloat: mark fl::basic_string + fl::sstream dtors noexcept to eliminate per-call-site landing pads #2975 Mark fl::basic_string + fl::sstream destructors noexcept to eliminate per-call-site landing pad metadata
• bloat: move more FL_WARN/FL_ERROR per-call-site code into log_emit (shrink call-site burst from ~200 B to ~50 B) #2982 Move more of the inline FL_WARN/FL_ERROR per-call-site burst into log_emit (shrink call-site cost from ~200 B to ~50 B)
• bloat: inline-storage fast path in fl::basic_string::write (skip variant dispatch for the common case) #2976 Inline-storage fast path in fl::basic_string::write(char const*, unsigned int)

Per-function:

• bloat: consolidate error log sites in ChannelEngineRMTImpl::reconfigureForNetwork + createChannel through expected<>::failure #2983 fl::ChannelEngineRMTImpl::reconfigureForNetwork() (2,378 B) + createChannel(...) (2,198 B) — consolidate error paths through single log site via expected<>::failure
• bloat: virtualize per-chipset writer dispatch in fl::Channel::showPixels (stop inlining 3 writer templates into one symbol) #2981 fl::Channel::showPixels(PixelController<RGB,1,-1>&) (1,884 B) — virtualize dynamic driver dispatch to stop inlining multiple writer templates
• bloat: extract addDriver validation + log paths into [[gnu::cold]] helper #2979 fl::ChannelManager::addDriver(int, shared_ptr<IChannelDriver>) (1,756 B) — extract validation/log paths into [[gnu::cold]] helper
• bloat: extract RmtMemoryManager::handleAllocateTxFailure format machinery into shared cold helper #2978 fl::RmtMemoryManager::handleAllocateTxFailure(...) (1,710 B) — extract format-and-log machinery into shared cold helper
• bloat: consolidate log sites in Rmt5EncoderImpl::initialize + extract esp-idf-error-wrap helper #2977 fl::detail::Rmt5EncoderImpl::initialize(...) (1,466 B) — consolidate multiple log sites + extract encoder-builder error wrapping
• bloat: reduce dispatch-loop landing pad count in ChannelEngineRMTImpl::processPendingChannels #2980 fl::ChannelEngineRMTImpl::processPendingChannels() (1,078 B) — reduce dispatch-loop landing pad count

Expected combined savings

If the cross-cutting fixes (A + B + C) land and at least 4 of the per-function fixes ship, we should recover ~6-8 KB of flash on ESP32-S3 Blink — roughly 2% of total flash, but more importantly it removes the pattern that makes every FastLED function pay the logging tax twice (call site + cleanup metadata).

Investigation tooling

• Bloat report: .build/symbols/esp32s3-fbuild/report.md (3,428 symbols)
• Back-reference graphviz sidecars: .build/symbols/esp32s3-fbuild/graphs/0001..0269.dot
• Per-symbol analysis script: ci/tmp/fl_bloat_investigation.py
• Per-function detail: .build/fl_bloat_investigation.md

Refs

• fl::sstream cost-reduction audit: flip the default so FL_WARN/FL_ERROR are compiled by default #2963 (gamma/sstream audit + Option 3 log_emit context)
• Meta: ESP32-S3 Blink bloat reduction (file:line concrete plan) — supersedes #2885 #2886 (the original ESP32-S3 NEOPIXEL bloat audit)
• perf(log): default FASTLED_LOG_VERBOSITY=0 in release builds (#2886 Stage 1) #2890 (FASTLED_LOG_VERBOSITY release-mode gating — already merged, this issue is the next step)