[Revisit 2] Project Hermes Revisited: An Erbium Host Chosen with Angie's Mixes, Two Routes to Light, and an Honest Account of What Light Can and Cannot Do on a Power Wafer

Project Hermes Revisited asks the next question after Angelica Revisited: can the same rare-earth-oxide GaN-on-Si foundation also be a light source at 1.54 µm, the wavelength of optical fibre? With the same reproducible, sourced multiphysics engine, it adds erbium rate equations (with concentration quenching and up-conversion), a leaky-mode optical solver for the GaN/oxide/Si stack, a pump-wavelength study, an amplifier model, electro-thermal coupling to the HEMT, a quantum-memory linewidth estimate and an optical-link energy budget. Every claim is tagged IMPLEMENTABLE, TESTABLE or CONCEPT, and one command regenerates every number. Five answers. (1) Host: Angie's Mixes, extended with an energy-transfer filter, picks Gd₂O₃:Er. It is optically silent and within 0.20 % of Si at growth, whereas the first edition's pure Er₂O₃ is 2.66 % off and quenches itself. (2) Foundation: putting Er into the Angelica base changes GaN stress by at most 0.08 MPa and bow by at most 0.6 µm, whether the oxide stays strained or relaxes. The designs used are those of Angelica Revisited v4, with physical superlattices. (3) Light in the foundation (Route L) can be made but not guided: only 0.0005 of the optical mode reaches the oxide. Its best use is a built-in temperature sensor under the transistor. (4) Light in the GaN (Route E): a GaN:Er layer 0.5 µm below the channel overlaps the mode by 0.43, at less than 0.01 % mobility cost. On today's best GaN waveguide loss it would give up to 6.4 dB/cm net gain; on the first GaN:Er waveguides it would not, so the open question is how much loss erbium adds. (5) Speed: erbium's millisecond lifetime caps direct modulation at about 10⁵ bit/s, so Er is a light source, amplifier, sensor or quantum memory, not a fast switch. An engineering hand-off turns this into a first lot: three test vehicles, five ridge-guide designs with pump-depleted gain, fibre coupling, ring designs, saturation and noise figure, growth and contamination checks, and a decision tree for every failed test. The first edition's audio chapter is removed, with the reason given. Formal paper in English (authoritative), Traditional Chinese and Japanese; primers for middle-school readers in the same three languages; full code and sources. Not peer-reviewed. Defensive publication. Text CC BY 4.0, code Apache 2.0. Relation to the first edition (June 2026, DOI 10.5281/zenodo.20621436): this is a rewrite. The pure Er₂O₃ layer of the first edition is replaced by dilute Er in a Gd₂O₃ host, the audio application is withdrawn, and every number is recomputed on the Angelica Revisited foundation.